Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
4 * All Rights Reserved.
5 */
6#include <linux/iversion.h>
7
8#include "xfs.h"
9#include "xfs_fs.h"
10#include "xfs_shared.h"
11#include "xfs_format.h"
12#include "xfs_log_format.h"
13#include "xfs_trans_resv.h"
14#include "xfs_sb.h"
15#include "xfs_mount.h"
16#include "xfs_defer.h"
17#include "xfs_inode.h"
18#include "xfs_dir2.h"
19#include "xfs_attr.h"
20#include "xfs_trans_space.h"
21#include "xfs_trans.h"
22#include "xfs_buf_item.h"
23#include "xfs_inode_item.h"
24#include "xfs_ialloc.h"
25#include "xfs_bmap.h"
26#include "xfs_bmap_util.h"
27#include "xfs_errortag.h"
28#include "xfs_error.h"
29#include "xfs_quota.h"
30#include "xfs_filestream.h"
31#include "xfs_trace.h"
32#include "xfs_icache.h"
33#include "xfs_symlink.h"
34#include "xfs_trans_priv.h"
35#include "xfs_log.h"
36#include "xfs_bmap_btree.h"
37#include "xfs_reflink.h"
38
39kmem_zone_t *xfs_inode_zone;
40
41/*
42 * Used in xfs_itruncate_extents(). This is the maximum number of extents
43 * freed from a file in a single transaction.
44 */
45#define XFS_ITRUNC_MAX_EXTENTS 2
46
47STATIC int xfs_iflush_int(struct xfs_inode *, struct xfs_buf *);
48STATIC int xfs_iunlink(struct xfs_trans *, struct xfs_inode *);
49STATIC int xfs_iunlink_remove(struct xfs_trans *, struct xfs_inode *);
50
51/*
52 * helper function to extract extent size hint from inode
53 */
54xfs_extlen_t
55xfs_get_extsz_hint(
56 struct xfs_inode *ip)
57{
58 if ((ip->i_d.di_flags & XFS_DIFLAG_EXTSIZE) && ip->i_d.di_extsize)
59 return ip->i_d.di_extsize;
60 if (XFS_IS_REALTIME_INODE(ip))
61 return ip->i_mount->m_sb.sb_rextsize;
62 return 0;
63}
64
65/*
66 * Helper function to extract CoW extent size hint from inode.
67 * Between the extent size hint and the CoW extent size hint, we
68 * return the greater of the two. If the value is zero (automatic),
69 * use the default size.
70 */
71xfs_extlen_t
72xfs_get_cowextsz_hint(
73 struct xfs_inode *ip)
74{
75 xfs_extlen_t a, b;
76
77 a = 0;
78 if (ip->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE)
79 a = ip->i_d.di_cowextsize;
80 b = xfs_get_extsz_hint(ip);
81
82 a = max(a, b);
83 if (a == 0)
84 return XFS_DEFAULT_COWEXTSZ_HINT;
85 return a;
86}
87
88/*
89 * These two are wrapper routines around the xfs_ilock() routine used to
90 * centralize some grungy code. They are used in places that wish to lock the
91 * inode solely for reading the extents. The reason these places can't just
92 * call xfs_ilock(ip, XFS_ILOCK_SHARED) is that the inode lock also guards to
93 * bringing in of the extents from disk for a file in b-tree format. If the
94 * inode is in b-tree format, then we need to lock the inode exclusively until
95 * the extents are read in. Locking it exclusively all the time would limit
96 * our parallelism unnecessarily, though. What we do instead is check to see
97 * if the extents have been read in yet, and only lock the inode exclusively
98 * if they have not.
99 *
100 * The functions return a value which should be given to the corresponding
101 * xfs_iunlock() call.
102 */
103uint
104xfs_ilock_data_map_shared(
105 struct xfs_inode *ip)
106{
107 uint lock_mode = XFS_ILOCK_SHARED;
108
109 if (ip->i_d.di_format == XFS_DINODE_FMT_BTREE &&
110 (ip->i_df.if_flags & XFS_IFEXTENTS) == 0)
111 lock_mode = XFS_ILOCK_EXCL;
112 xfs_ilock(ip, lock_mode);
113 return lock_mode;
114}
115
116uint
117xfs_ilock_attr_map_shared(
118 struct xfs_inode *ip)
119{
120 uint lock_mode = XFS_ILOCK_SHARED;
121
122 if (ip->i_d.di_aformat == XFS_DINODE_FMT_BTREE &&
123 (ip->i_afp->if_flags & XFS_IFEXTENTS) == 0)
124 lock_mode = XFS_ILOCK_EXCL;
125 xfs_ilock(ip, lock_mode);
126 return lock_mode;
127}
128
129/*
130 * In addition to i_rwsem in the VFS inode, the xfs inode contains 2
131 * multi-reader locks: i_mmap_lock and the i_lock. This routine allows
132 * various combinations of the locks to be obtained.
133 *
134 * The 3 locks should always be ordered so that the IO lock is obtained first,
135 * the mmap lock second and the ilock last in order to prevent deadlock.
136 *
137 * Basic locking order:
138 *
139 * i_rwsem -> i_mmap_lock -> page_lock -> i_ilock
140 *
141 * mmap_sem locking order:
142 *
143 * i_rwsem -> page lock -> mmap_sem
144 * mmap_sem -> i_mmap_lock -> page_lock
145 *
146 * The difference in mmap_sem locking order mean that we cannot hold the
147 * i_mmap_lock over syscall based read(2)/write(2) based IO. These IO paths can
148 * fault in pages during copy in/out (for buffered IO) or require the mmap_sem
149 * in get_user_pages() to map the user pages into the kernel address space for
150 * direct IO. Similarly the i_rwsem cannot be taken inside a page fault because
151 * page faults already hold the mmap_sem.
152 *
153 * Hence to serialise fully against both syscall and mmap based IO, we need to
154 * take both the i_rwsem and the i_mmap_lock. These locks should *only* be both
155 * taken in places where we need to invalidate the page cache in a race
156 * free manner (e.g. truncate, hole punch and other extent manipulation
157 * functions).
158 */
159void
160xfs_ilock(
161 xfs_inode_t *ip,
162 uint lock_flags)
163{
164 trace_xfs_ilock(ip, lock_flags, _RET_IP_);
165
166 /*
167 * You can't set both SHARED and EXCL for the same lock,
168 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
169 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
170 */
171 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
172 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
173 ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
174 (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
175 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
176 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
177 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
178
179 if (lock_flags & XFS_IOLOCK_EXCL) {
180 down_write_nested(&VFS_I(ip)->i_rwsem,
181 XFS_IOLOCK_DEP(lock_flags));
182 } else if (lock_flags & XFS_IOLOCK_SHARED) {
183 down_read_nested(&VFS_I(ip)->i_rwsem,
184 XFS_IOLOCK_DEP(lock_flags));
185 }
186
187 if (lock_flags & XFS_MMAPLOCK_EXCL)
188 mrupdate_nested(&ip->i_mmaplock, XFS_MMAPLOCK_DEP(lock_flags));
189 else if (lock_flags & XFS_MMAPLOCK_SHARED)
190 mraccess_nested(&ip->i_mmaplock, XFS_MMAPLOCK_DEP(lock_flags));
191
192 if (lock_flags & XFS_ILOCK_EXCL)
193 mrupdate_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
194 else if (lock_flags & XFS_ILOCK_SHARED)
195 mraccess_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
196}
197
198/*
199 * This is just like xfs_ilock(), except that the caller
200 * is guaranteed not to sleep. It returns 1 if it gets
201 * the requested locks and 0 otherwise. If the IO lock is
202 * obtained but the inode lock cannot be, then the IO lock
203 * is dropped before returning.
204 *
205 * ip -- the inode being locked
206 * lock_flags -- this parameter indicates the inode's locks to be
207 * to be locked. See the comment for xfs_ilock() for a list
208 * of valid values.
209 */
210int
211xfs_ilock_nowait(
212 xfs_inode_t *ip,
213 uint lock_flags)
214{
215 trace_xfs_ilock_nowait(ip, lock_flags, _RET_IP_);
216
217 /*
218 * You can't set both SHARED and EXCL for the same lock,
219 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
220 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
221 */
222 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
223 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
224 ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
225 (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
226 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
227 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
228 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
229
230 if (lock_flags & XFS_IOLOCK_EXCL) {
231 if (!down_write_trylock(&VFS_I(ip)->i_rwsem))
232 goto out;
233 } else if (lock_flags & XFS_IOLOCK_SHARED) {
234 if (!down_read_trylock(&VFS_I(ip)->i_rwsem))
235 goto out;
236 }
237
238 if (lock_flags & XFS_MMAPLOCK_EXCL) {
239 if (!mrtryupdate(&ip->i_mmaplock))
240 goto out_undo_iolock;
241 } else if (lock_flags & XFS_MMAPLOCK_SHARED) {
242 if (!mrtryaccess(&ip->i_mmaplock))
243 goto out_undo_iolock;
244 }
245
246 if (lock_flags & XFS_ILOCK_EXCL) {
247 if (!mrtryupdate(&ip->i_lock))
248 goto out_undo_mmaplock;
249 } else if (lock_flags & XFS_ILOCK_SHARED) {
250 if (!mrtryaccess(&ip->i_lock))
251 goto out_undo_mmaplock;
252 }
253 return 1;
254
255out_undo_mmaplock:
256 if (lock_flags & XFS_MMAPLOCK_EXCL)
257 mrunlock_excl(&ip->i_mmaplock);
258 else if (lock_flags & XFS_MMAPLOCK_SHARED)
259 mrunlock_shared(&ip->i_mmaplock);
260out_undo_iolock:
261 if (lock_flags & XFS_IOLOCK_EXCL)
262 up_write(&VFS_I(ip)->i_rwsem);
263 else if (lock_flags & XFS_IOLOCK_SHARED)
264 up_read(&VFS_I(ip)->i_rwsem);
265out:
266 return 0;
267}
268
269/*
270 * xfs_iunlock() is used to drop the inode locks acquired with
271 * xfs_ilock() and xfs_ilock_nowait(). The caller must pass
272 * in the flags given to xfs_ilock() or xfs_ilock_nowait() so
273 * that we know which locks to drop.
274 *
275 * ip -- the inode being unlocked
276 * lock_flags -- this parameter indicates the inode's locks to be
277 * to be unlocked. See the comment for xfs_ilock() for a list
278 * of valid values for this parameter.
279 *
280 */
281void
282xfs_iunlock(
283 xfs_inode_t *ip,
284 uint lock_flags)
285{
286 /*
287 * You can't set both SHARED and EXCL for the same lock,
288 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
289 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
290 */
291 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
292 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
293 ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
294 (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
295 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
296 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
297 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
298 ASSERT(lock_flags != 0);
299
300 if (lock_flags & XFS_IOLOCK_EXCL)
301 up_write(&VFS_I(ip)->i_rwsem);
302 else if (lock_flags & XFS_IOLOCK_SHARED)
303 up_read(&VFS_I(ip)->i_rwsem);
304
305 if (lock_flags & XFS_MMAPLOCK_EXCL)
306 mrunlock_excl(&ip->i_mmaplock);
307 else if (lock_flags & XFS_MMAPLOCK_SHARED)
308 mrunlock_shared(&ip->i_mmaplock);
309
310 if (lock_flags & XFS_ILOCK_EXCL)
311 mrunlock_excl(&ip->i_lock);
312 else if (lock_flags & XFS_ILOCK_SHARED)
313 mrunlock_shared(&ip->i_lock);
314
315 trace_xfs_iunlock(ip, lock_flags, _RET_IP_);
316}
317
318/*
319 * give up write locks. the i/o lock cannot be held nested
320 * if it is being demoted.
321 */
322void
323xfs_ilock_demote(
324 xfs_inode_t *ip,
325 uint lock_flags)
326{
327 ASSERT(lock_flags & (XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL));
328 ASSERT((lock_flags &
329 ~(XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL)) == 0);
330
331 if (lock_flags & XFS_ILOCK_EXCL)
332 mrdemote(&ip->i_lock);
333 if (lock_flags & XFS_MMAPLOCK_EXCL)
334 mrdemote(&ip->i_mmaplock);
335 if (lock_flags & XFS_IOLOCK_EXCL)
336 downgrade_write(&VFS_I(ip)->i_rwsem);
337
338 trace_xfs_ilock_demote(ip, lock_flags, _RET_IP_);
339}
340
341#if defined(DEBUG) || defined(XFS_WARN)
342int
343xfs_isilocked(
344 xfs_inode_t *ip,
345 uint lock_flags)
346{
347 if (lock_flags & (XFS_ILOCK_EXCL|XFS_ILOCK_SHARED)) {
348 if (!(lock_flags & XFS_ILOCK_SHARED))
349 return !!ip->i_lock.mr_writer;
350 return rwsem_is_locked(&ip->i_lock.mr_lock);
351 }
352
353 if (lock_flags & (XFS_MMAPLOCK_EXCL|XFS_MMAPLOCK_SHARED)) {
354 if (!(lock_flags & XFS_MMAPLOCK_SHARED))
355 return !!ip->i_mmaplock.mr_writer;
356 return rwsem_is_locked(&ip->i_mmaplock.mr_lock);
357 }
358
359 if (lock_flags & (XFS_IOLOCK_EXCL|XFS_IOLOCK_SHARED)) {
360 if (!(lock_flags & XFS_IOLOCK_SHARED))
361 return !debug_locks ||
362 lockdep_is_held_type(&VFS_I(ip)->i_rwsem, 0);
363 return rwsem_is_locked(&VFS_I(ip)->i_rwsem);
364 }
365
366 ASSERT(0);
367 return 0;
368}
369#endif
370
371/*
372 * xfs_lockdep_subclass_ok() is only used in an ASSERT, so is only called when
373 * DEBUG or XFS_WARN is set. And MAX_LOCKDEP_SUBCLASSES is then only defined
374 * when CONFIG_LOCKDEP is set. Hence the complex define below to avoid build
375 * errors and warnings.
376 */
377#if (defined(DEBUG) || defined(XFS_WARN)) && defined(CONFIG_LOCKDEP)
378static bool
379xfs_lockdep_subclass_ok(
380 int subclass)
381{
382 return subclass < MAX_LOCKDEP_SUBCLASSES;
383}
384#else
385#define xfs_lockdep_subclass_ok(subclass) (true)
386#endif
387
388/*
389 * Bump the subclass so xfs_lock_inodes() acquires each lock with a different
390 * value. This can be called for any type of inode lock combination, including
391 * parent locking. Care must be taken to ensure we don't overrun the subclass
392 * storage fields in the class mask we build.
393 */
394static inline int
395xfs_lock_inumorder(int lock_mode, int subclass)
396{
397 int class = 0;
398
399 ASSERT(!(lock_mode & (XFS_ILOCK_PARENT | XFS_ILOCK_RTBITMAP |
400 XFS_ILOCK_RTSUM)));
401 ASSERT(xfs_lockdep_subclass_ok(subclass));
402
403 if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)) {
404 ASSERT(subclass <= XFS_IOLOCK_MAX_SUBCLASS);
405 class += subclass << XFS_IOLOCK_SHIFT;
406 }
407
408 if (lock_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) {
409 ASSERT(subclass <= XFS_MMAPLOCK_MAX_SUBCLASS);
410 class += subclass << XFS_MMAPLOCK_SHIFT;
411 }
412
413 if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) {
414 ASSERT(subclass <= XFS_ILOCK_MAX_SUBCLASS);
415 class += subclass << XFS_ILOCK_SHIFT;
416 }
417
418 return (lock_mode & ~XFS_LOCK_SUBCLASS_MASK) | class;
419}
420
421/*
422 * The following routine will lock n inodes in exclusive mode. We assume the
423 * caller calls us with the inodes in i_ino order.
424 *
425 * We need to detect deadlock where an inode that we lock is in the AIL and we
426 * start waiting for another inode that is locked by a thread in a long running
427 * transaction (such as truncate). This can result in deadlock since the long
428 * running trans might need to wait for the inode we just locked in order to
429 * push the tail and free space in the log.
430 *
431 * xfs_lock_inodes() can only be used to lock one type of lock at a time -
432 * the iolock, the mmaplock or the ilock, but not more than one at a time. If we
433 * lock more than one at a time, lockdep will report false positives saying we
434 * have violated locking orders.
435 */
436static void
437xfs_lock_inodes(
438 struct xfs_inode **ips,
439 int inodes,
440 uint lock_mode)
441{
442 int attempts = 0, i, j, try_lock;
443 struct xfs_log_item *lp;
444
445 /*
446 * Currently supports between 2 and 5 inodes with exclusive locking. We
447 * support an arbitrary depth of locking here, but absolute limits on
448 * inodes depend on the the type of locking and the limits placed by
449 * lockdep annotations in xfs_lock_inumorder. These are all checked by
450 * the asserts.
451 */
452 ASSERT(ips && inodes >= 2 && inodes <= 5);
453 ASSERT(lock_mode & (XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL |
454 XFS_ILOCK_EXCL));
455 ASSERT(!(lock_mode & (XFS_IOLOCK_SHARED | XFS_MMAPLOCK_SHARED |
456 XFS_ILOCK_SHARED)));
457 ASSERT(!(lock_mode & XFS_MMAPLOCK_EXCL) ||
458 inodes <= XFS_MMAPLOCK_MAX_SUBCLASS + 1);
459 ASSERT(!(lock_mode & XFS_ILOCK_EXCL) ||
460 inodes <= XFS_ILOCK_MAX_SUBCLASS + 1);
461
462 if (lock_mode & XFS_IOLOCK_EXCL) {
463 ASSERT(!(lock_mode & (XFS_MMAPLOCK_EXCL | XFS_ILOCK_EXCL)));
464 } else if (lock_mode & XFS_MMAPLOCK_EXCL)
465 ASSERT(!(lock_mode & XFS_ILOCK_EXCL));
466
467 try_lock = 0;
468 i = 0;
469again:
470 for (; i < inodes; i++) {
471 ASSERT(ips[i]);
472
473 if (i && (ips[i] == ips[i - 1])) /* Already locked */
474 continue;
475
476 /*
477 * If try_lock is not set yet, make sure all locked inodes are
478 * not in the AIL. If any are, set try_lock to be used later.
479 */
480 if (!try_lock) {
481 for (j = (i - 1); j >= 0 && !try_lock; j--) {
482 lp = &ips[j]->i_itemp->ili_item;
483 if (lp && test_bit(XFS_LI_IN_AIL, &lp->li_flags))
484 try_lock++;
485 }
486 }
487
488 /*
489 * If any of the previous locks we have locked is in the AIL,
490 * we must TRY to get the second and subsequent locks. If
491 * we can't get any, we must release all we have
492 * and try again.
493 */
494 if (!try_lock) {
495 xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
496 continue;
497 }
498
499 /* try_lock means we have an inode locked that is in the AIL. */
500 ASSERT(i != 0);
501 if (xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i)))
502 continue;
503
504 /*
505 * Unlock all previous guys and try again. xfs_iunlock will try
506 * to push the tail if the inode is in the AIL.
507 */
508 attempts++;
509 for (j = i - 1; j >= 0; j--) {
510 /*
511 * Check to see if we've already unlocked this one. Not
512 * the first one going back, and the inode ptr is the
513 * same.
514 */
515 if (j != (i - 1) && ips[j] == ips[j + 1])
516 continue;
517
518 xfs_iunlock(ips[j], lock_mode);
519 }
520
521 if ((attempts % 5) == 0) {
522 delay(1); /* Don't just spin the CPU */
523 }
524 i = 0;
525 try_lock = 0;
526 goto again;
527 }
528}
529
530/*
531 * xfs_lock_two_inodes() can only be used to lock one type of lock at a time -
532 * the mmaplock or the ilock, but not more than one type at a time. If we lock
533 * more than one at a time, lockdep will report false positives saying we have
534 * violated locking orders. The iolock must be double-locked separately since
535 * we use i_rwsem for that. We now support taking one lock EXCL and the other
536 * SHARED.
537 */
538void
539xfs_lock_two_inodes(
540 struct xfs_inode *ip0,
541 uint ip0_mode,
542 struct xfs_inode *ip1,
543 uint ip1_mode)
544{
545 struct xfs_inode *temp;
546 uint mode_temp;
547 int attempts = 0;
548 struct xfs_log_item *lp;
549
550 ASSERT(hweight32(ip0_mode) == 1);
551 ASSERT(hweight32(ip1_mode) == 1);
552 ASSERT(!(ip0_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
553 ASSERT(!(ip1_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
554 ASSERT(!(ip0_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) ||
555 !(ip0_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
556 ASSERT(!(ip1_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) ||
557 !(ip1_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
558 ASSERT(!(ip1_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) ||
559 !(ip0_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
560 ASSERT(!(ip0_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) ||
561 !(ip1_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
562
563 ASSERT(ip0->i_ino != ip1->i_ino);
564
565 if (ip0->i_ino > ip1->i_ino) {
566 temp = ip0;
567 ip0 = ip1;
568 ip1 = temp;
569 mode_temp = ip0_mode;
570 ip0_mode = ip1_mode;
571 ip1_mode = mode_temp;
572 }
573
574 again:
575 xfs_ilock(ip0, xfs_lock_inumorder(ip0_mode, 0));
576
577 /*
578 * If the first lock we have locked is in the AIL, we must TRY to get
579 * the second lock. If we can't get it, we must release the first one
580 * and try again.
581 */
582 lp = &ip0->i_itemp->ili_item;
583 if (lp && test_bit(XFS_LI_IN_AIL, &lp->li_flags)) {
584 if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(ip1_mode, 1))) {
585 xfs_iunlock(ip0, ip0_mode);
586 if ((++attempts % 5) == 0)
587 delay(1); /* Don't just spin the CPU */
588 goto again;
589 }
590 } else {
591 xfs_ilock(ip1, xfs_lock_inumorder(ip1_mode, 1));
592 }
593}
594
595void
596__xfs_iflock(
597 struct xfs_inode *ip)
598{
599 wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_IFLOCK_BIT);
600 DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IFLOCK_BIT);
601
602 do {
603 prepare_to_wait_exclusive(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
604 if (xfs_isiflocked(ip))
605 io_schedule();
606 } while (!xfs_iflock_nowait(ip));
607
608 finish_wait(wq, &wait.wq_entry);
609}
610
611STATIC uint
612_xfs_dic2xflags(
613 uint16_t di_flags,
614 uint64_t di_flags2,
615 bool has_attr)
616{
617 uint flags = 0;
618
619 if (di_flags & XFS_DIFLAG_ANY) {
620 if (di_flags & XFS_DIFLAG_REALTIME)
621 flags |= FS_XFLAG_REALTIME;
622 if (di_flags & XFS_DIFLAG_PREALLOC)
623 flags |= FS_XFLAG_PREALLOC;
624 if (di_flags & XFS_DIFLAG_IMMUTABLE)
625 flags |= FS_XFLAG_IMMUTABLE;
626 if (di_flags & XFS_DIFLAG_APPEND)
627 flags |= FS_XFLAG_APPEND;
628 if (di_flags & XFS_DIFLAG_SYNC)
629 flags |= FS_XFLAG_SYNC;
630 if (di_flags & XFS_DIFLAG_NOATIME)
631 flags |= FS_XFLAG_NOATIME;
632 if (di_flags & XFS_DIFLAG_NODUMP)
633 flags |= FS_XFLAG_NODUMP;
634 if (di_flags & XFS_DIFLAG_RTINHERIT)
635 flags |= FS_XFLAG_RTINHERIT;
636 if (di_flags & XFS_DIFLAG_PROJINHERIT)
637 flags |= FS_XFLAG_PROJINHERIT;
638 if (di_flags & XFS_DIFLAG_NOSYMLINKS)
639 flags |= FS_XFLAG_NOSYMLINKS;
640 if (di_flags & XFS_DIFLAG_EXTSIZE)
641 flags |= FS_XFLAG_EXTSIZE;
642 if (di_flags & XFS_DIFLAG_EXTSZINHERIT)
643 flags |= FS_XFLAG_EXTSZINHERIT;
644 if (di_flags & XFS_DIFLAG_NODEFRAG)
645 flags |= FS_XFLAG_NODEFRAG;
646 if (di_flags & XFS_DIFLAG_FILESTREAM)
647 flags |= FS_XFLAG_FILESTREAM;
648 }
649
650 if (di_flags2 & XFS_DIFLAG2_ANY) {
651 if (di_flags2 & XFS_DIFLAG2_DAX)
652 flags |= FS_XFLAG_DAX;
653 if (di_flags2 & XFS_DIFLAG2_COWEXTSIZE)
654 flags |= FS_XFLAG_COWEXTSIZE;
655 }
656
657 if (has_attr)
658 flags |= FS_XFLAG_HASATTR;
659
660 return flags;
661}
662
663uint
664xfs_ip2xflags(
665 struct xfs_inode *ip)
666{
667 struct xfs_icdinode *dic = &ip->i_d;
668
669 return _xfs_dic2xflags(dic->di_flags, dic->di_flags2, XFS_IFORK_Q(ip));
670}
671
672/*
673 * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
674 * is allowed, otherwise it has to be an exact match. If a CI match is found,
675 * ci_name->name will point to a the actual name (caller must free) or
676 * will be set to NULL if an exact match is found.
677 */
678int
679xfs_lookup(
680 xfs_inode_t *dp,
681 struct xfs_name *name,
682 xfs_inode_t **ipp,
683 struct xfs_name *ci_name)
684{
685 xfs_ino_t inum;
686 int error;
687
688 trace_xfs_lookup(dp, name);
689
690 if (XFS_FORCED_SHUTDOWN(dp->i_mount))
691 return -EIO;
692
693 error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
694 if (error)
695 goto out_unlock;
696
697 error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp);
698 if (error)
699 goto out_free_name;
700
701 return 0;
702
703out_free_name:
704 if (ci_name)
705 kmem_free(ci_name->name);
706out_unlock:
707 *ipp = NULL;
708 return error;
709}
710
711/*
712 * Allocate an inode on disk and return a copy of its in-core version.
713 * The in-core inode is locked exclusively. Set mode, nlink, and rdev
714 * appropriately within the inode. The uid and gid for the inode are
715 * set according to the contents of the given cred structure.
716 *
717 * Use xfs_dialloc() to allocate the on-disk inode. If xfs_dialloc()
718 * has a free inode available, call xfs_iget() to obtain the in-core
719 * version of the allocated inode. Finally, fill in the inode and
720 * log its initial contents. In this case, ialloc_context would be
721 * set to NULL.
722 *
723 * If xfs_dialloc() does not have an available inode, it will replenish
724 * its supply by doing an allocation. Since we can only do one
725 * allocation within a transaction without deadlocks, we must commit
726 * the current transaction before returning the inode itself.
727 * In this case, therefore, we will set ialloc_context and return.
728 * The caller should then commit the current transaction, start a new
729 * transaction, and call xfs_ialloc() again to actually get the inode.
730 *
731 * To ensure that some other process does not grab the inode that
732 * was allocated during the first call to xfs_ialloc(), this routine
733 * also returns the [locked] bp pointing to the head of the freelist
734 * as ialloc_context. The caller should hold this buffer across
735 * the commit and pass it back into this routine on the second call.
736 *
737 * If we are allocating quota inodes, we do not have a parent inode
738 * to attach to or associate with (i.e. pip == NULL) because they
739 * are not linked into the directory structure - they are attached
740 * directly to the superblock - and so have no parent.
741 */
742static int
743xfs_ialloc(
744 xfs_trans_t *tp,
745 xfs_inode_t *pip,
746 umode_t mode,
747 xfs_nlink_t nlink,
748 dev_t rdev,
749 prid_t prid,
750 xfs_buf_t **ialloc_context,
751 xfs_inode_t **ipp)
752{
753 struct xfs_mount *mp = tp->t_mountp;
754 xfs_ino_t ino;
755 xfs_inode_t *ip;
756 uint flags;
757 int error;
758 struct timespec64 tv;
759 struct inode *inode;
760
761 /*
762 * Call the space management code to pick
763 * the on-disk inode to be allocated.
764 */
765 error = xfs_dialloc(tp, pip ? pip->i_ino : 0, mode,
766 ialloc_context, &ino);
767 if (error)
768 return error;
769 if (*ialloc_context || ino == NULLFSINO) {
770 *ipp = NULL;
771 return 0;
772 }
773 ASSERT(*ialloc_context == NULL);
774
775 /*
776 * Protect against obviously corrupt allocation btree records. Later
777 * xfs_iget checks will catch re-allocation of other active in-memory
778 * and on-disk inodes. If we don't catch reallocating the parent inode
779 * here we will deadlock in xfs_iget() so we have to do these checks
780 * first.
781 */
782 if ((pip && ino == pip->i_ino) || !xfs_verify_dir_ino(mp, ino)) {
783 xfs_alert(mp, "Allocated a known in-use inode 0x%llx!", ino);
784 return -EFSCORRUPTED;
785 }
786
787 /*
788 * Get the in-core inode with the lock held exclusively.
789 * This is because we're setting fields here we need
790 * to prevent others from looking at until we're done.
791 */
792 error = xfs_iget(mp, tp, ino, XFS_IGET_CREATE,
793 XFS_ILOCK_EXCL, &ip);
794 if (error)
795 return error;
796 ASSERT(ip != NULL);
797 inode = VFS_I(ip);
798
799 /*
800 * We always convert v1 inodes to v2 now - we only support filesystems
801 * with >= v2 inode capability, so there is no reason for ever leaving
802 * an inode in v1 format.
803 */
804 if (ip->i_d.di_version == 1)
805 ip->i_d.di_version = 2;
806
807 inode->i_mode = mode;
808 set_nlink(inode, nlink);
809 ip->i_d.di_uid = xfs_kuid_to_uid(current_fsuid());
810 ip->i_d.di_gid = xfs_kgid_to_gid(current_fsgid());
811 inode->i_rdev = rdev;
812 xfs_set_projid(ip, prid);
813
814 if (pip && XFS_INHERIT_GID(pip)) {
815 ip->i_d.di_gid = pip->i_d.di_gid;
816 if ((VFS_I(pip)->i_mode & S_ISGID) && S_ISDIR(mode))
817 inode->i_mode |= S_ISGID;
818 }
819
820 /*
821 * If the group ID of the new file does not match the effective group
822 * ID or one of the supplementary group IDs, the S_ISGID bit is cleared
823 * (and only if the irix_sgid_inherit compatibility variable is set).
824 */
825 if ((irix_sgid_inherit) &&
826 (inode->i_mode & S_ISGID) &&
827 (!in_group_p(xfs_gid_to_kgid(ip->i_d.di_gid))))
828 inode->i_mode &= ~S_ISGID;
829
830 ip->i_d.di_size = 0;
831 ip->i_d.di_nextents = 0;
832 ASSERT(ip->i_d.di_nblocks == 0);
833
834 tv = current_time(inode);
835 inode->i_mtime = tv;
836 inode->i_atime = tv;
837 inode->i_ctime = tv;
838
839 ip->i_d.di_extsize = 0;
840 ip->i_d.di_dmevmask = 0;
841 ip->i_d.di_dmstate = 0;
842 ip->i_d.di_flags = 0;
843
844 if (ip->i_d.di_version == 3) {
845 inode_set_iversion(inode, 1);
846 ip->i_d.di_flags2 = 0;
847 ip->i_d.di_cowextsize = 0;
848 ip->i_d.di_crtime.t_sec = (int32_t)tv.tv_sec;
849 ip->i_d.di_crtime.t_nsec = (int32_t)tv.tv_nsec;
850 }
851
852
853 flags = XFS_ILOG_CORE;
854 switch (mode & S_IFMT) {
855 case S_IFIFO:
856 case S_IFCHR:
857 case S_IFBLK:
858 case S_IFSOCK:
859 ip->i_d.di_format = XFS_DINODE_FMT_DEV;
860 ip->i_df.if_flags = 0;
861 flags |= XFS_ILOG_DEV;
862 break;
863 case S_IFREG:
864 case S_IFDIR:
865 if (pip && (pip->i_d.di_flags & XFS_DIFLAG_ANY)) {
866 uint di_flags = 0;
867
868 if (S_ISDIR(mode)) {
869 if (pip->i_d.di_flags & XFS_DIFLAG_RTINHERIT)
870 di_flags |= XFS_DIFLAG_RTINHERIT;
871 if (pip->i_d.di_flags & XFS_DIFLAG_EXTSZINHERIT) {
872 di_flags |= XFS_DIFLAG_EXTSZINHERIT;
873 ip->i_d.di_extsize = pip->i_d.di_extsize;
874 }
875 if (pip->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
876 di_flags |= XFS_DIFLAG_PROJINHERIT;
877 } else if (S_ISREG(mode)) {
878 if (pip->i_d.di_flags & XFS_DIFLAG_RTINHERIT)
879 di_flags |= XFS_DIFLAG_REALTIME;
880 if (pip->i_d.di_flags & XFS_DIFLAG_EXTSZINHERIT) {
881 di_flags |= XFS_DIFLAG_EXTSIZE;
882 ip->i_d.di_extsize = pip->i_d.di_extsize;
883 }
884 }
885 if ((pip->i_d.di_flags & XFS_DIFLAG_NOATIME) &&
886 xfs_inherit_noatime)
887 di_flags |= XFS_DIFLAG_NOATIME;
888 if ((pip->i_d.di_flags & XFS_DIFLAG_NODUMP) &&
889 xfs_inherit_nodump)
890 di_flags |= XFS_DIFLAG_NODUMP;
891 if ((pip->i_d.di_flags & XFS_DIFLAG_SYNC) &&
892 xfs_inherit_sync)
893 di_flags |= XFS_DIFLAG_SYNC;
894 if ((pip->i_d.di_flags & XFS_DIFLAG_NOSYMLINKS) &&
895 xfs_inherit_nosymlinks)
896 di_flags |= XFS_DIFLAG_NOSYMLINKS;
897 if ((pip->i_d.di_flags & XFS_DIFLAG_NODEFRAG) &&
898 xfs_inherit_nodefrag)
899 di_flags |= XFS_DIFLAG_NODEFRAG;
900 if (pip->i_d.di_flags & XFS_DIFLAG_FILESTREAM)
901 di_flags |= XFS_DIFLAG_FILESTREAM;
902
903 ip->i_d.di_flags |= di_flags;
904 }
905 if (pip &&
906 (pip->i_d.di_flags2 & XFS_DIFLAG2_ANY) &&
907 pip->i_d.di_version == 3 &&
908 ip->i_d.di_version == 3) {
909 uint64_t di_flags2 = 0;
910
911 if (pip->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE) {
912 di_flags2 |= XFS_DIFLAG2_COWEXTSIZE;
913 ip->i_d.di_cowextsize = pip->i_d.di_cowextsize;
914 }
915 if (pip->i_d.di_flags2 & XFS_DIFLAG2_DAX)
916 di_flags2 |= XFS_DIFLAG2_DAX;
917
918 ip->i_d.di_flags2 |= di_flags2;
919 }
920 /* FALLTHROUGH */
921 case S_IFLNK:
922 ip->i_d.di_format = XFS_DINODE_FMT_EXTENTS;
923 ip->i_df.if_flags = XFS_IFEXTENTS;
924 ip->i_df.if_bytes = 0;
925 ip->i_df.if_u1.if_root = NULL;
926 break;
927 default:
928 ASSERT(0);
929 }
930 /*
931 * Attribute fork settings for new inode.
932 */
933 ip->i_d.di_aformat = XFS_DINODE_FMT_EXTENTS;
934 ip->i_d.di_anextents = 0;
935
936 /*
937 * Log the new values stuffed into the inode.
938 */
939 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
940 xfs_trans_log_inode(tp, ip, flags);
941
942 /* now that we have an i_mode we can setup the inode structure */
943 xfs_setup_inode(ip);
944
945 *ipp = ip;
946 return 0;
947}
948
949/*
950 * Allocates a new inode from disk and return a pointer to the
951 * incore copy. This routine will internally commit the current
952 * transaction and allocate a new one if the Space Manager needed
953 * to do an allocation to replenish the inode free-list.
954 *
955 * This routine is designed to be called from xfs_create and
956 * xfs_create_dir.
957 *
958 */
959int
960xfs_dir_ialloc(
961 xfs_trans_t **tpp, /* input: current transaction;
962 output: may be a new transaction. */
963 xfs_inode_t *dp, /* directory within whose allocate
964 the inode. */
965 umode_t mode,
966 xfs_nlink_t nlink,
967 dev_t rdev,
968 prid_t prid, /* project id */
969 xfs_inode_t **ipp) /* pointer to inode; it will be
970 locked. */
971{
972 xfs_trans_t *tp;
973 xfs_inode_t *ip;
974 xfs_buf_t *ialloc_context = NULL;
975 int code;
976 void *dqinfo;
977 uint tflags;
978
979 tp = *tpp;
980 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
981
982 /*
983 * xfs_ialloc will return a pointer to an incore inode if
984 * the Space Manager has an available inode on the free
985 * list. Otherwise, it will do an allocation and replenish
986 * the freelist. Since we can only do one allocation per
987 * transaction without deadlocks, we will need to commit the
988 * current transaction and start a new one. We will then
989 * need to call xfs_ialloc again to get the inode.
990 *
991 * If xfs_ialloc did an allocation to replenish the freelist,
992 * it returns the bp containing the head of the freelist as
993 * ialloc_context. We will hold a lock on it across the
994 * transaction commit so that no other process can steal
995 * the inode(s) that we've just allocated.
996 */
997 code = xfs_ialloc(tp, dp, mode, nlink, rdev, prid, &ialloc_context,
998 &ip);
999
1000 /*
1001 * Return an error if we were unable to allocate a new inode.
1002 * This should only happen if we run out of space on disk or
1003 * encounter a disk error.
1004 */
1005 if (code) {
1006 *ipp = NULL;
1007 return code;
1008 }
1009 if (!ialloc_context && !ip) {
1010 *ipp = NULL;
1011 return -ENOSPC;
1012 }
1013
1014 /*
1015 * If the AGI buffer is non-NULL, then we were unable to get an
1016 * inode in one operation. We need to commit the current
1017 * transaction and call xfs_ialloc() again. It is guaranteed
1018 * to succeed the second time.
1019 */
1020 if (ialloc_context) {
1021 /*
1022 * Normally, xfs_trans_commit releases all the locks.
1023 * We call bhold to hang on to the ialloc_context across
1024 * the commit. Holding this buffer prevents any other
1025 * processes from doing any allocations in this
1026 * allocation group.
1027 */
1028 xfs_trans_bhold(tp, ialloc_context);
1029
1030 /*
1031 * We want the quota changes to be associated with the next
1032 * transaction, NOT this one. So, detach the dqinfo from this
1033 * and attach it to the next transaction.
1034 */
1035 dqinfo = NULL;
1036 tflags = 0;
1037 if (tp->t_dqinfo) {
1038 dqinfo = (void *)tp->t_dqinfo;
1039 tp->t_dqinfo = NULL;
1040 tflags = tp->t_flags & XFS_TRANS_DQ_DIRTY;
1041 tp->t_flags &= ~(XFS_TRANS_DQ_DIRTY);
1042 }
1043
1044 code = xfs_trans_roll(&tp);
1045
1046 /*
1047 * Re-attach the quota info that we detached from prev trx.
1048 */
1049 if (dqinfo) {
1050 tp->t_dqinfo = dqinfo;
1051 tp->t_flags |= tflags;
1052 }
1053
1054 if (code) {
1055 xfs_buf_relse(ialloc_context);
1056 *tpp = tp;
1057 *ipp = NULL;
1058 return code;
1059 }
1060 xfs_trans_bjoin(tp, ialloc_context);
1061
1062 /*
1063 * Call ialloc again. Since we've locked out all
1064 * other allocations in this allocation group,
1065 * this call should always succeed.
1066 */
1067 code = xfs_ialloc(tp, dp, mode, nlink, rdev, prid,
1068 &ialloc_context, &ip);
1069
1070 /*
1071 * If we get an error at this point, return to the caller
1072 * so that the current transaction can be aborted.
1073 */
1074 if (code) {
1075 *tpp = tp;
1076 *ipp = NULL;
1077 return code;
1078 }
1079 ASSERT(!ialloc_context && ip);
1080
1081 }
1082
1083 *ipp = ip;
1084 *tpp = tp;
1085
1086 return 0;
1087}
1088
1089/*
1090 * Decrement the link count on an inode & log the change. If this causes the
1091 * link count to go to zero, move the inode to AGI unlinked list so that it can
1092 * be freed when the last active reference goes away via xfs_inactive().
1093 */
1094static int /* error */
1095xfs_droplink(
1096 xfs_trans_t *tp,
1097 xfs_inode_t *ip)
1098{
1099 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
1100
1101 drop_nlink(VFS_I(ip));
1102 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1103
1104 if (VFS_I(ip)->i_nlink)
1105 return 0;
1106
1107 return xfs_iunlink(tp, ip);
1108}
1109
1110/*
1111 * Increment the link count on an inode & log the change.
1112 */
1113static void
1114xfs_bumplink(
1115 xfs_trans_t *tp,
1116 xfs_inode_t *ip)
1117{
1118 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
1119
1120 ASSERT(ip->i_d.di_version > 1);
1121 inc_nlink(VFS_I(ip));
1122 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1123}
1124
1125int
1126xfs_create(
1127 xfs_inode_t *dp,
1128 struct xfs_name *name,
1129 umode_t mode,
1130 dev_t rdev,
1131 xfs_inode_t **ipp)
1132{
1133 int is_dir = S_ISDIR(mode);
1134 struct xfs_mount *mp = dp->i_mount;
1135 struct xfs_inode *ip = NULL;
1136 struct xfs_trans *tp = NULL;
1137 int error;
1138 bool unlock_dp_on_error = false;
1139 prid_t prid;
1140 struct xfs_dquot *udqp = NULL;
1141 struct xfs_dquot *gdqp = NULL;
1142 struct xfs_dquot *pdqp = NULL;
1143 struct xfs_trans_res *tres;
1144 uint resblks;
1145
1146 trace_xfs_create(dp, name);
1147
1148 if (XFS_FORCED_SHUTDOWN(mp))
1149 return -EIO;
1150
1151 prid = xfs_get_initial_prid(dp);
1152
1153 /*
1154 * Make sure that we have allocated dquot(s) on disk.
1155 */
1156 error = xfs_qm_vop_dqalloc(dp, xfs_kuid_to_uid(current_fsuid()),
1157 xfs_kgid_to_gid(current_fsgid()), prid,
1158 XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
1159 &udqp, &gdqp, &pdqp);
1160 if (error)
1161 return error;
1162
1163 if (is_dir) {
1164 resblks = XFS_MKDIR_SPACE_RES(mp, name->len);
1165 tres = &M_RES(mp)->tr_mkdir;
1166 } else {
1167 resblks = XFS_CREATE_SPACE_RES(mp, name->len);
1168 tres = &M_RES(mp)->tr_create;
1169 }
1170
1171 /*
1172 * Initially assume that the file does not exist and
1173 * reserve the resources for that case. If that is not
1174 * the case we'll drop the one we have and get a more
1175 * appropriate transaction later.
1176 */
1177 error = xfs_trans_alloc(mp, tres, resblks, 0, 0, &tp);
1178 if (error == -ENOSPC) {
1179 /* flush outstanding delalloc blocks and retry */
1180 xfs_flush_inodes(mp);
1181 error = xfs_trans_alloc(mp, tres, resblks, 0, 0, &tp);
1182 }
1183 if (error)
1184 goto out_release_inode;
1185
1186 xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
1187 unlock_dp_on_error = true;
1188
1189 /*
1190 * Reserve disk quota and the inode.
1191 */
1192 error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp,
1193 pdqp, resblks, 1, 0);
1194 if (error)
1195 goto out_trans_cancel;
1196
1197 /*
1198 * A newly created regular or special file just has one directory
1199 * entry pointing to them, but a directory also the "." entry
1200 * pointing to itself.
1201 */
1202 error = xfs_dir_ialloc(&tp, dp, mode, is_dir ? 2 : 1, rdev, prid, &ip);
1203 if (error)
1204 goto out_trans_cancel;
1205
1206 /*
1207 * Now we join the directory inode to the transaction. We do not do it
1208 * earlier because xfs_dir_ialloc might commit the previous transaction
1209 * (and release all the locks). An error from here on will result in
1210 * the transaction cancel unlocking dp so don't do it explicitly in the
1211 * error path.
1212 */
1213 xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1214 unlock_dp_on_error = false;
1215
1216 error = xfs_dir_createname(tp, dp, name, ip->i_ino,
1217 resblks ?
1218 resblks - XFS_IALLOC_SPACE_RES(mp) : 0);
1219 if (error) {
1220 ASSERT(error != -ENOSPC);
1221 goto out_trans_cancel;
1222 }
1223 xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1224 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1225
1226 if (is_dir) {
1227 error = xfs_dir_init(tp, ip, dp);
1228 if (error)
1229 goto out_trans_cancel;
1230
1231 xfs_bumplink(tp, dp);
1232 }
1233
1234 /*
1235 * If this is a synchronous mount, make sure that the
1236 * create transaction goes to disk before returning to
1237 * the user.
1238 */
1239 if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1240 xfs_trans_set_sync(tp);
1241
1242 /*
1243 * Attach the dquot(s) to the inodes and modify them incore.
1244 * These ids of the inode couldn't have changed since the new
1245 * inode has been locked ever since it was created.
1246 */
1247 xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1248
1249 error = xfs_trans_commit(tp);
1250 if (error)
1251 goto out_release_inode;
1252
1253 xfs_qm_dqrele(udqp);
1254 xfs_qm_dqrele(gdqp);
1255 xfs_qm_dqrele(pdqp);
1256
1257 *ipp = ip;
1258 return 0;
1259
1260 out_trans_cancel:
1261 xfs_trans_cancel(tp);
1262 out_release_inode:
1263 /*
1264 * Wait until after the current transaction is aborted to finish the
1265 * setup of the inode and release the inode. This prevents recursive
1266 * transactions and deadlocks from xfs_inactive.
1267 */
1268 if (ip) {
1269 xfs_finish_inode_setup(ip);
1270 xfs_irele(ip);
1271 }
1272
1273 xfs_qm_dqrele(udqp);
1274 xfs_qm_dqrele(gdqp);
1275 xfs_qm_dqrele(pdqp);
1276
1277 if (unlock_dp_on_error)
1278 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1279 return error;
1280}
1281
1282int
1283xfs_create_tmpfile(
1284 struct xfs_inode *dp,
1285 umode_t mode,
1286 struct xfs_inode **ipp)
1287{
1288 struct xfs_mount *mp = dp->i_mount;
1289 struct xfs_inode *ip = NULL;
1290 struct xfs_trans *tp = NULL;
1291 int error;
1292 prid_t prid;
1293 struct xfs_dquot *udqp = NULL;
1294 struct xfs_dquot *gdqp = NULL;
1295 struct xfs_dquot *pdqp = NULL;
1296 struct xfs_trans_res *tres;
1297 uint resblks;
1298
1299 if (XFS_FORCED_SHUTDOWN(mp))
1300 return -EIO;
1301
1302 prid = xfs_get_initial_prid(dp);
1303
1304 /*
1305 * Make sure that we have allocated dquot(s) on disk.
1306 */
1307 error = xfs_qm_vop_dqalloc(dp, xfs_kuid_to_uid(current_fsuid()),
1308 xfs_kgid_to_gid(current_fsgid()), prid,
1309 XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
1310 &udqp, &gdqp, &pdqp);
1311 if (error)
1312 return error;
1313
1314 resblks = XFS_IALLOC_SPACE_RES(mp);
1315 tres = &M_RES(mp)->tr_create_tmpfile;
1316
1317 error = xfs_trans_alloc(mp, tres, resblks, 0, 0, &tp);
1318 if (error)
1319 goto out_release_inode;
1320
1321 error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp,
1322 pdqp, resblks, 1, 0);
1323 if (error)
1324 goto out_trans_cancel;
1325
1326 error = xfs_dir_ialloc(&tp, dp, mode, 0, 0, prid, &ip);
1327 if (error)
1328 goto out_trans_cancel;
1329
1330 if (mp->m_flags & XFS_MOUNT_WSYNC)
1331 xfs_trans_set_sync(tp);
1332
1333 /*
1334 * Attach the dquot(s) to the inodes and modify them incore.
1335 * These ids of the inode couldn't have changed since the new
1336 * inode has been locked ever since it was created.
1337 */
1338 xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1339
1340 error = xfs_iunlink(tp, ip);
1341 if (error)
1342 goto out_trans_cancel;
1343
1344 error = xfs_trans_commit(tp);
1345 if (error)
1346 goto out_release_inode;
1347
1348 xfs_qm_dqrele(udqp);
1349 xfs_qm_dqrele(gdqp);
1350 xfs_qm_dqrele(pdqp);
1351
1352 *ipp = ip;
1353 return 0;
1354
1355 out_trans_cancel:
1356 xfs_trans_cancel(tp);
1357 out_release_inode:
1358 /*
1359 * Wait until after the current transaction is aborted to finish the
1360 * setup of the inode and release the inode. This prevents recursive
1361 * transactions and deadlocks from xfs_inactive.
1362 */
1363 if (ip) {
1364 xfs_finish_inode_setup(ip);
1365 xfs_irele(ip);
1366 }
1367
1368 xfs_qm_dqrele(udqp);
1369 xfs_qm_dqrele(gdqp);
1370 xfs_qm_dqrele(pdqp);
1371
1372 return error;
1373}
1374
1375int
1376xfs_link(
1377 xfs_inode_t *tdp,
1378 xfs_inode_t *sip,
1379 struct xfs_name *target_name)
1380{
1381 xfs_mount_t *mp = tdp->i_mount;
1382 xfs_trans_t *tp;
1383 int error;
1384 int resblks;
1385
1386 trace_xfs_link(tdp, target_name);
1387
1388 ASSERT(!S_ISDIR(VFS_I(sip)->i_mode));
1389
1390 if (XFS_FORCED_SHUTDOWN(mp))
1391 return -EIO;
1392
1393 error = xfs_qm_dqattach(sip);
1394 if (error)
1395 goto std_return;
1396
1397 error = xfs_qm_dqattach(tdp);
1398 if (error)
1399 goto std_return;
1400
1401 resblks = XFS_LINK_SPACE_RES(mp, target_name->len);
1402 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_link, resblks, 0, 0, &tp);
1403 if (error == -ENOSPC) {
1404 resblks = 0;
1405 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_link, 0, 0, 0, &tp);
1406 }
1407 if (error)
1408 goto std_return;
1409
1410 xfs_lock_two_inodes(sip, XFS_ILOCK_EXCL, tdp, XFS_ILOCK_EXCL);
1411
1412 xfs_trans_ijoin(tp, sip, XFS_ILOCK_EXCL);
1413 xfs_trans_ijoin(tp, tdp, XFS_ILOCK_EXCL);
1414
1415 /*
1416 * If we are using project inheritance, we only allow hard link
1417 * creation in our tree when the project IDs are the same; else
1418 * the tree quota mechanism could be circumvented.
1419 */
1420 if (unlikely((tdp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
1421 (xfs_get_projid(tdp) != xfs_get_projid(sip)))) {
1422 error = -EXDEV;
1423 goto error_return;
1424 }
1425
1426 if (!resblks) {
1427 error = xfs_dir_canenter(tp, tdp, target_name);
1428 if (error)
1429 goto error_return;
1430 }
1431
1432 /*
1433 * Handle initial link state of O_TMPFILE inode
1434 */
1435 if (VFS_I(sip)->i_nlink == 0) {
1436 error = xfs_iunlink_remove(tp, sip);
1437 if (error)
1438 goto error_return;
1439 }
1440
1441 error = xfs_dir_createname(tp, tdp, target_name, sip->i_ino,
1442 resblks);
1443 if (error)
1444 goto error_return;
1445 xfs_trans_ichgtime(tp, tdp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1446 xfs_trans_log_inode(tp, tdp, XFS_ILOG_CORE);
1447
1448 xfs_bumplink(tp, sip);
1449
1450 /*
1451 * If this is a synchronous mount, make sure that the
1452 * link transaction goes to disk before returning to
1453 * the user.
1454 */
1455 if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1456 xfs_trans_set_sync(tp);
1457
1458 return xfs_trans_commit(tp);
1459
1460 error_return:
1461 xfs_trans_cancel(tp);
1462 std_return:
1463 return error;
1464}
1465
1466/* Clear the reflink flag and the cowblocks tag if possible. */
1467static void
1468xfs_itruncate_clear_reflink_flags(
1469 struct xfs_inode *ip)
1470{
1471 struct xfs_ifork *dfork;
1472 struct xfs_ifork *cfork;
1473
1474 if (!xfs_is_reflink_inode(ip))
1475 return;
1476 dfork = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1477 cfork = XFS_IFORK_PTR(ip, XFS_COW_FORK);
1478 if (dfork->if_bytes == 0 && cfork->if_bytes == 0)
1479 ip->i_d.di_flags2 &= ~XFS_DIFLAG2_REFLINK;
1480 if (cfork->if_bytes == 0)
1481 xfs_inode_clear_cowblocks_tag(ip);
1482}
1483
1484/*
1485 * Free up the underlying blocks past new_size. The new size must be smaller
1486 * than the current size. This routine can be used both for the attribute and
1487 * data fork, and does not modify the inode size, which is left to the caller.
1488 *
1489 * The transaction passed to this routine must have made a permanent log
1490 * reservation of at least XFS_ITRUNCATE_LOG_RES. This routine may commit the
1491 * given transaction and start new ones, so make sure everything involved in
1492 * the transaction is tidy before calling here. Some transaction will be
1493 * returned to the caller to be committed. The incoming transaction must
1494 * already include the inode, and both inode locks must be held exclusively.
1495 * The inode must also be "held" within the transaction. On return the inode
1496 * will be "held" within the returned transaction. This routine does NOT
1497 * require any disk space to be reserved for it within the transaction.
1498 *
1499 * If we get an error, we must return with the inode locked and linked into the
1500 * current transaction. This keeps things simple for the higher level code,
1501 * because it always knows that the inode is locked and held in the transaction
1502 * that returns to it whether errors occur or not. We don't mark the inode
1503 * dirty on error so that transactions can be easily aborted if possible.
1504 */
1505int
1506xfs_itruncate_extents_flags(
1507 struct xfs_trans **tpp,
1508 struct xfs_inode *ip,
1509 int whichfork,
1510 xfs_fsize_t new_size,
1511 int flags)
1512{
1513 struct xfs_mount *mp = ip->i_mount;
1514 struct xfs_trans *tp = *tpp;
1515 xfs_fileoff_t first_unmap_block;
1516 xfs_fileoff_t last_block;
1517 xfs_filblks_t unmap_len;
1518 int error = 0;
1519 int done = 0;
1520
1521 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1522 ASSERT(!atomic_read(&VFS_I(ip)->i_count) ||
1523 xfs_isilocked(ip, XFS_IOLOCK_EXCL));
1524 ASSERT(new_size <= XFS_ISIZE(ip));
1525 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
1526 ASSERT(ip->i_itemp != NULL);
1527 ASSERT(ip->i_itemp->ili_lock_flags == 0);
1528 ASSERT(!XFS_NOT_DQATTACHED(mp, ip));
1529
1530 trace_xfs_itruncate_extents_start(ip, new_size);
1531
1532 flags |= xfs_bmapi_aflag(whichfork);
1533
1534 /*
1535 * Since it is possible for space to become allocated beyond
1536 * the end of the file (in a crash where the space is allocated
1537 * but the inode size is not yet updated), simply remove any
1538 * blocks which show up between the new EOF and the maximum
1539 * possible file size. If the first block to be removed is
1540 * beyond the maximum file size (ie it is the same as last_block),
1541 * then there is nothing to do.
1542 */
1543 first_unmap_block = XFS_B_TO_FSB(mp, (xfs_ufsize_t)new_size);
1544 last_block = XFS_B_TO_FSB(mp, mp->m_super->s_maxbytes);
1545 if (first_unmap_block == last_block)
1546 return 0;
1547
1548 ASSERT(first_unmap_block < last_block);
1549 unmap_len = last_block - first_unmap_block + 1;
1550 while (!done) {
1551 ASSERT(tp->t_firstblock == NULLFSBLOCK);
1552 error = xfs_bunmapi(tp, ip, first_unmap_block, unmap_len, flags,
1553 XFS_ITRUNC_MAX_EXTENTS, &done);
1554 if (error)
1555 goto out;
1556
1557 /*
1558 * Duplicate the transaction that has the permanent
1559 * reservation and commit the old transaction.
1560 */
1561 error = xfs_defer_finish(&tp);
1562 if (error)
1563 goto out;
1564
1565 error = xfs_trans_roll_inode(&tp, ip);
1566 if (error)
1567 goto out;
1568 }
1569
1570 if (whichfork == XFS_DATA_FORK) {
1571 /* Remove all pending CoW reservations. */
1572 error = xfs_reflink_cancel_cow_blocks(ip, &tp,
1573 first_unmap_block, last_block, true);
1574 if (error)
1575 goto out;
1576
1577 xfs_itruncate_clear_reflink_flags(ip);
1578 }
1579
1580 /*
1581 * Always re-log the inode so that our permanent transaction can keep
1582 * on rolling it forward in the log.
1583 */
1584 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1585
1586 trace_xfs_itruncate_extents_end(ip, new_size);
1587
1588out:
1589 *tpp = tp;
1590 return error;
1591}
1592
1593int
1594xfs_release(
1595 xfs_inode_t *ip)
1596{
1597 xfs_mount_t *mp = ip->i_mount;
1598 int error;
1599
1600 if (!S_ISREG(VFS_I(ip)->i_mode) || (VFS_I(ip)->i_mode == 0))
1601 return 0;
1602
1603 /* If this is a read-only mount, don't do this (would generate I/O) */
1604 if (mp->m_flags & XFS_MOUNT_RDONLY)
1605 return 0;
1606
1607 if (!XFS_FORCED_SHUTDOWN(mp)) {
1608 int truncated;
1609
1610 /*
1611 * If we previously truncated this file and removed old data
1612 * in the process, we want to initiate "early" writeout on
1613 * the last close. This is an attempt to combat the notorious
1614 * NULL files problem which is particularly noticeable from a
1615 * truncate down, buffered (re-)write (delalloc), followed by
1616 * a crash. What we are effectively doing here is
1617 * significantly reducing the time window where we'd otherwise
1618 * be exposed to that problem.
1619 */
1620 truncated = xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED);
1621 if (truncated) {
1622 xfs_iflags_clear(ip, XFS_IDIRTY_RELEASE);
1623 if (ip->i_delayed_blks > 0) {
1624 error = filemap_flush(VFS_I(ip)->i_mapping);
1625 if (error)
1626 return error;
1627 }
1628 }
1629 }
1630
1631 if (VFS_I(ip)->i_nlink == 0)
1632 return 0;
1633
1634 if (xfs_can_free_eofblocks(ip, false)) {
1635
1636 /*
1637 * Check if the inode is being opened, written and closed
1638 * frequently and we have delayed allocation blocks outstanding
1639 * (e.g. streaming writes from the NFS server), truncating the
1640 * blocks past EOF will cause fragmentation to occur.
1641 *
1642 * In this case don't do the truncation, but we have to be
1643 * careful how we detect this case. Blocks beyond EOF show up as
1644 * i_delayed_blks even when the inode is clean, so we need to
1645 * truncate them away first before checking for a dirty release.
1646 * Hence on the first dirty close we will still remove the
1647 * speculative allocation, but after that we will leave it in
1648 * place.
1649 */
1650 if (xfs_iflags_test(ip, XFS_IDIRTY_RELEASE))
1651 return 0;
1652 /*
1653 * If we can't get the iolock just skip truncating the blocks
1654 * past EOF because we could deadlock with the mmap_sem
1655 * otherwise. We'll get another chance to drop them once the
1656 * last reference to the inode is dropped, so we'll never leak
1657 * blocks permanently.
1658 */
1659 if (xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1660 error = xfs_free_eofblocks(ip);
1661 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1662 if (error)
1663 return error;
1664 }
1665
1666 /* delalloc blocks after truncation means it really is dirty */
1667 if (ip->i_delayed_blks)
1668 xfs_iflags_set(ip, XFS_IDIRTY_RELEASE);
1669 }
1670 return 0;
1671}
1672
1673/*
1674 * xfs_inactive_truncate
1675 *
1676 * Called to perform a truncate when an inode becomes unlinked.
1677 */
1678STATIC int
1679xfs_inactive_truncate(
1680 struct xfs_inode *ip)
1681{
1682 struct xfs_mount *mp = ip->i_mount;
1683 struct xfs_trans *tp;
1684 int error;
1685
1686 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate, 0, 0, 0, &tp);
1687 if (error) {
1688 ASSERT(XFS_FORCED_SHUTDOWN(mp));
1689 return error;
1690 }
1691 xfs_ilock(ip, XFS_ILOCK_EXCL);
1692 xfs_trans_ijoin(tp, ip, 0);
1693
1694 /*
1695 * Log the inode size first to prevent stale data exposure in the event
1696 * of a system crash before the truncate completes. See the related
1697 * comment in xfs_vn_setattr_size() for details.
1698 */
1699 ip->i_d.di_size = 0;
1700 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1701
1702 error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK, 0);
1703 if (error)
1704 goto error_trans_cancel;
1705
1706 ASSERT(ip->i_d.di_nextents == 0);
1707
1708 error = xfs_trans_commit(tp);
1709 if (error)
1710 goto error_unlock;
1711
1712 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1713 return 0;
1714
1715error_trans_cancel:
1716 xfs_trans_cancel(tp);
1717error_unlock:
1718 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1719 return error;
1720}
1721
1722/*
1723 * xfs_inactive_ifree()
1724 *
1725 * Perform the inode free when an inode is unlinked.
1726 */
1727STATIC int
1728xfs_inactive_ifree(
1729 struct xfs_inode *ip)
1730{
1731 struct xfs_mount *mp = ip->i_mount;
1732 struct xfs_trans *tp;
1733 int error;
1734
1735 /*
1736 * We try to use a per-AG reservation for any block needed by the finobt
1737 * tree, but as the finobt feature predates the per-AG reservation
1738 * support a degraded file system might not have enough space for the
1739 * reservation at mount time. In that case try to dip into the reserved
1740 * pool and pray.
1741 *
1742 * Send a warning if the reservation does happen to fail, as the inode
1743 * now remains allocated and sits on the unlinked list until the fs is
1744 * repaired.
1745 */
1746 if (unlikely(mp->m_finobt_nores)) {
1747 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree,
1748 XFS_IFREE_SPACE_RES(mp), 0, XFS_TRANS_RESERVE,
1749 &tp);
1750 } else {
1751 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree, 0, 0, 0, &tp);
1752 }
1753 if (error) {
1754 if (error == -ENOSPC) {
1755 xfs_warn_ratelimited(mp,
1756 "Failed to remove inode(s) from unlinked list. "
1757 "Please free space, unmount and run xfs_repair.");
1758 } else {
1759 ASSERT(XFS_FORCED_SHUTDOWN(mp));
1760 }
1761 return error;
1762 }
1763
1764 xfs_ilock(ip, XFS_ILOCK_EXCL);
1765 xfs_trans_ijoin(tp, ip, 0);
1766
1767 error = xfs_ifree(tp, ip);
1768 if (error) {
1769 /*
1770 * If we fail to free the inode, shut down. The cancel
1771 * might do that, we need to make sure. Otherwise the
1772 * inode might be lost for a long time or forever.
1773 */
1774 if (!XFS_FORCED_SHUTDOWN(mp)) {
1775 xfs_notice(mp, "%s: xfs_ifree returned error %d",
1776 __func__, error);
1777 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1778 }
1779 xfs_trans_cancel(tp);
1780 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1781 return error;
1782 }
1783
1784 /*
1785 * Credit the quota account(s). The inode is gone.
1786 */
1787 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
1788
1789 /*
1790 * Just ignore errors at this point. There is nothing we can do except
1791 * to try to keep going. Make sure it's not a silent error.
1792 */
1793 error = xfs_trans_commit(tp);
1794 if (error)
1795 xfs_notice(mp, "%s: xfs_trans_commit returned error %d",
1796 __func__, error);
1797
1798 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1799 return 0;
1800}
1801
1802/*
1803 * xfs_inactive
1804 *
1805 * This is called when the vnode reference count for the vnode
1806 * goes to zero. If the file has been unlinked, then it must
1807 * now be truncated. Also, we clear all of the read-ahead state
1808 * kept for the inode here since the file is now closed.
1809 */
1810void
1811xfs_inactive(
1812 xfs_inode_t *ip)
1813{
1814 struct xfs_mount *mp;
1815 int error;
1816 int truncate = 0;
1817
1818 /*
1819 * If the inode is already free, then there can be nothing
1820 * to clean up here.
1821 */
1822 if (VFS_I(ip)->i_mode == 0) {
1823 ASSERT(ip->i_df.if_broot_bytes == 0);
1824 return;
1825 }
1826
1827 mp = ip->i_mount;
1828 ASSERT(!xfs_iflags_test(ip, XFS_IRECOVERY));
1829
1830 /* If this is a read-only mount, don't do this (would generate I/O) */
1831 if (mp->m_flags & XFS_MOUNT_RDONLY)
1832 return;
1833
1834 /* Try to clean out the cow blocks if there are any. */
1835 if (xfs_inode_has_cow_data(ip))
1836 xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, true);
1837
1838 if (VFS_I(ip)->i_nlink != 0) {
1839 /*
1840 * force is true because we are evicting an inode from the
1841 * cache. Post-eof blocks must be freed, lest we end up with
1842 * broken free space accounting.
1843 *
1844 * Note: don't bother with iolock here since lockdep complains
1845 * about acquiring it in reclaim context. We have the only
1846 * reference to the inode at this point anyways.
1847 */
1848 if (xfs_can_free_eofblocks(ip, true))
1849 xfs_free_eofblocks(ip);
1850
1851 return;
1852 }
1853
1854 if (S_ISREG(VFS_I(ip)->i_mode) &&
1855 (ip->i_d.di_size != 0 || XFS_ISIZE(ip) != 0 ||
1856 ip->i_d.di_nextents > 0 || ip->i_delayed_blks > 0))
1857 truncate = 1;
1858
1859 error = xfs_qm_dqattach(ip);
1860 if (error)
1861 return;
1862
1863 if (S_ISLNK(VFS_I(ip)->i_mode))
1864 error = xfs_inactive_symlink(ip);
1865 else if (truncate)
1866 error = xfs_inactive_truncate(ip);
1867 if (error)
1868 return;
1869
1870 /*
1871 * If there are attributes associated with the file then blow them away
1872 * now. The code calls a routine that recursively deconstructs the
1873 * attribute fork. If also blows away the in-core attribute fork.
1874 */
1875 if (XFS_IFORK_Q(ip)) {
1876 error = xfs_attr_inactive(ip);
1877 if (error)
1878 return;
1879 }
1880
1881 ASSERT(!ip->i_afp);
1882 ASSERT(ip->i_d.di_anextents == 0);
1883 ASSERT(ip->i_d.di_forkoff == 0);
1884
1885 /*
1886 * Free the inode.
1887 */
1888 error = xfs_inactive_ifree(ip);
1889 if (error)
1890 return;
1891
1892 /*
1893 * Release the dquots held by inode, if any.
1894 */
1895 xfs_qm_dqdetach(ip);
1896}
1897
1898/*
1899 * In-Core Unlinked List Lookups
1900 * =============================
1901 *
1902 * Every inode is supposed to be reachable from some other piece of metadata
1903 * with the exception of the root directory. Inodes with a connection to a
1904 * file descriptor but not linked from anywhere in the on-disk directory tree
1905 * are collectively known as unlinked inodes, though the filesystem itself
1906 * maintains links to these inodes so that on-disk metadata are consistent.
1907 *
1908 * XFS implements a per-AG on-disk hash table of unlinked inodes. The AGI
1909 * header contains a number of buckets that point to an inode, and each inode
1910 * record has a pointer to the next inode in the hash chain. This
1911 * singly-linked list causes scaling problems in the iunlink remove function
1912 * because we must walk that list to find the inode that points to the inode
1913 * being removed from the unlinked hash bucket list.
1914 *
1915 * What if we modelled the unlinked list as a collection of records capturing
1916 * "X.next_unlinked = Y" relations? If we indexed those records on Y, we'd
1917 * have a fast way to look up unlinked list predecessors, which avoids the
1918 * slow list walk. That's exactly what we do here (in-core) with a per-AG
1919 * rhashtable.
1920 *
1921 * Because this is a backref cache, we ignore operational failures since the
1922 * iunlink code can fall back to the slow bucket walk. The only errors that
1923 * should bubble out are for obviously incorrect situations.
1924 *
1925 * All users of the backref cache MUST hold the AGI buffer lock to serialize
1926 * access or have otherwise provided for concurrency control.
1927 */
1928
1929/* Capture a "X.next_unlinked = Y" relationship. */
1930struct xfs_iunlink {
1931 struct rhash_head iu_rhash_head;
1932 xfs_agino_t iu_agino; /* X */
1933 xfs_agino_t iu_next_unlinked; /* Y */
1934};
1935
1936/* Unlinked list predecessor lookup hashtable construction */
1937static int
1938xfs_iunlink_obj_cmpfn(
1939 struct rhashtable_compare_arg *arg,
1940 const void *obj)
1941{
1942 const xfs_agino_t *key = arg->key;
1943 const struct xfs_iunlink *iu = obj;
1944
1945 if (iu->iu_next_unlinked != *key)
1946 return 1;
1947 return 0;
1948}
1949
1950static const struct rhashtable_params xfs_iunlink_hash_params = {
1951 .min_size = XFS_AGI_UNLINKED_BUCKETS,
1952 .key_len = sizeof(xfs_agino_t),
1953 .key_offset = offsetof(struct xfs_iunlink,
1954 iu_next_unlinked),
1955 .head_offset = offsetof(struct xfs_iunlink, iu_rhash_head),
1956 .automatic_shrinking = true,
1957 .obj_cmpfn = xfs_iunlink_obj_cmpfn,
1958};
1959
1960/*
1961 * Return X, where X.next_unlinked == @agino. Returns NULLAGINO if no such
1962 * relation is found.
1963 */
1964static xfs_agino_t
1965xfs_iunlink_lookup_backref(
1966 struct xfs_perag *pag,
1967 xfs_agino_t agino)
1968{
1969 struct xfs_iunlink *iu;
1970
1971 iu = rhashtable_lookup_fast(&pag->pagi_unlinked_hash, &agino,
1972 xfs_iunlink_hash_params);
1973 return iu ? iu->iu_agino : NULLAGINO;
1974}
1975
1976/*
1977 * Take ownership of an iunlink cache entry and insert it into the hash table.
1978 * If successful, the entry will be owned by the cache; if not, it is freed.
1979 * Either way, the caller does not own @iu after this call.
1980 */
1981static int
1982xfs_iunlink_insert_backref(
1983 struct xfs_perag *pag,
1984 struct xfs_iunlink *iu)
1985{
1986 int error;
1987
1988 error = rhashtable_insert_fast(&pag->pagi_unlinked_hash,
1989 &iu->iu_rhash_head, xfs_iunlink_hash_params);
1990 /*
1991 * Fail loudly if there already was an entry because that's a sign of
1992 * corruption of in-memory data. Also fail loudly if we see an error
1993 * code we didn't anticipate from the rhashtable code. Currently we
1994 * only anticipate ENOMEM.
1995 */
1996 if (error) {
1997 WARN(error != -ENOMEM, "iunlink cache insert error %d", error);
1998 kmem_free(iu);
1999 }
2000 /*
2001 * Absorb any runtime errors that aren't a result of corruption because
2002 * this is a cache and we can always fall back to bucket list scanning.
2003 */
2004 if (error != 0 && error != -EEXIST)
2005 error = 0;
2006 return error;
2007}
2008
2009/* Remember that @prev_agino.next_unlinked = @this_agino. */
2010static int
2011xfs_iunlink_add_backref(
2012 struct xfs_perag *pag,
2013 xfs_agino_t prev_agino,
2014 xfs_agino_t this_agino)
2015{
2016 struct xfs_iunlink *iu;
2017
2018 if (XFS_TEST_ERROR(false, pag->pag_mount, XFS_ERRTAG_IUNLINK_FALLBACK))
2019 return 0;
2020
2021 iu = kmem_zalloc(sizeof(*iu), KM_NOFS);
2022 iu->iu_agino = prev_agino;
2023 iu->iu_next_unlinked = this_agino;
2024
2025 return xfs_iunlink_insert_backref(pag, iu);
2026}
2027
2028/*
2029 * Replace X.next_unlinked = @agino with X.next_unlinked = @next_unlinked.
2030 * If @next_unlinked is NULLAGINO, we drop the backref and exit. If there
2031 * wasn't any such entry then we don't bother.
2032 */
2033static int
2034xfs_iunlink_change_backref(
2035 struct xfs_perag *pag,
2036 xfs_agino_t agino,
2037 xfs_agino_t next_unlinked)
2038{
2039 struct xfs_iunlink *iu;
2040 int error;
2041
2042 /* Look up the old entry; if there wasn't one then exit. */
2043 iu = rhashtable_lookup_fast(&pag->pagi_unlinked_hash, &agino,
2044 xfs_iunlink_hash_params);
2045 if (!iu)
2046 return 0;
2047
2048 /*
2049 * Remove the entry. This shouldn't ever return an error, but if we
2050 * couldn't remove the old entry we don't want to add it again to the
2051 * hash table, and if the entry disappeared on us then someone's
2052 * violated the locking rules and we need to fail loudly. Either way
2053 * we cannot remove the inode because internal state is or would have
2054 * been corrupt.
2055 */
2056 error = rhashtable_remove_fast(&pag->pagi_unlinked_hash,
2057 &iu->iu_rhash_head, xfs_iunlink_hash_params);
2058 if (error)
2059 return error;
2060
2061 /* If there is no new next entry just free our item and return. */
2062 if (next_unlinked == NULLAGINO) {
2063 kmem_free(iu);
2064 return 0;
2065 }
2066
2067 /* Update the entry and re-add it to the hash table. */
2068 iu->iu_next_unlinked = next_unlinked;
2069 return xfs_iunlink_insert_backref(pag, iu);
2070}
2071
2072/* Set up the in-core predecessor structures. */
2073int
2074xfs_iunlink_init(
2075 struct xfs_perag *pag)
2076{
2077 return rhashtable_init(&pag->pagi_unlinked_hash,
2078 &xfs_iunlink_hash_params);
2079}
2080
2081/* Free the in-core predecessor structures. */
2082static void
2083xfs_iunlink_free_item(
2084 void *ptr,
2085 void *arg)
2086{
2087 struct xfs_iunlink *iu = ptr;
2088 bool *freed_anything = arg;
2089
2090 *freed_anything = true;
2091 kmem_free(iu);
2092}
2093
2094void
2095xfs_iunlink_destroy(
2096 struct xfs_perag *pag)
2097{
2098 bool freed_anything = false;
2099
2100 rhashtable_free_and_destroy(&pag->pagi_unlinked_hash,
2101 xfs_iunlink_free_item, &freed_anything);
2102
2103 ASSERT(freed_anything == false || XFS_FORCED_SHUTDOWN(pag->pag_mount));
2104}
2105
2106/*
2107 * Point the AGI unlinked bucket at an inode and log the results. The caller
2108 * is responsible for validating the old value.
2109 */
2110STATIC int
2111xfs_iunlink_update_bucket(
2112 struct xfs_trans *tp,
2113 xfs_agnumber_t agno,
2114 struct xfs_buf *agibp,
2115 unsigned int bucket_index,
2116 xfs_agino_t new_agino)
2117{
2118 struct xfs_agi *agi = XFS_BUF_TO_AGI(agibp);
2119 xfs_agino_t old_value;
2120 int offset;
2121
2122 ASSERT(xfs_verify_agino_or_null(tp->t_mountp, agno, new_agino));
2123
2124 old_value = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2125 trace_xfs_iunlink_update_bucket(tp->t_mountp, agno, bucket_index,
2126 old_value, new_agino);
2127
2128 /*
2129 * We should never find the head of the list already set to the value
2130 * passed in because either we're adding or removing ourselves from the
2131 * head of the list.
2132 */
2133 if (old_value == new_agino)
2134 return -EFSCORRUPTED;
2135
2136 agi->agi_unlinked[bucket_index] = cpu_to_be32(new_agino);
2137 offset = offsetof(struct xfs_agi, agi_unlinked) +
2138 (sizeof(xfs_agino_t) * bucket_index);
2139 xfs_trans_log_buf(tp, agibp, offset, offset + sizeof(xfs_agino_t) - 1);
2140 return 0;
2141}
2142
2143/* Set an on-disk inode's next_unlinked pointer. */
2144STATIC void
2145xfs_iunlink_update_dinode(
2146 struct xfs_trans *tp,
2147 xfs_agnumber_t agno,
2148 xfs_agino_t agino,
2149 struct xfs_buf *ibp,
2150 struct xfs_dinode *dip,
2151 struct xfs_imap *imap,
2152 xfs_agino_t next_agino)
2153{
2154 struct xfs_mount *mp = tp->t_mountp;
2155 int offset;
2156
2157 ASSERT(xfs_verify_agino_or_null(mp, agno, next_agino));
2158
2159 trace_xfs_iunlink_update_dinode(mp, agno, agino,
2160 be32_to_cpu(dip->di_next_unlinked), next_agino);
2161
2162 dip->di_next_unlinked = cpu_to_be32(next_agino);
2163 offset = imap->im_boffset +
2164 offsetof(struct xfs_dinode, di_next_unlinked);
2165
2166 /* need to recalc the inode CRC if appropriate */
2167 xfs_dinode_calc_crc(mp, dip);
2168 xfs_trans_inode_buf(tp, ibp);
2169 xfs_trans_log_buf(tp, ibp, offset, offset + sizeof(xfs_agino_t) - 1);
2170 xfs_inobp_check(mp, ibp);
2171}
2172
2173/* Set an in-core inode's unlinked pointer and return the old value. */
2174STATIC int
2175xfs_iunlink_update_inode(
2176 struct xfs_trans *tp,
2177 struct xfs_inode *ip,
2178 xfs_agnumber_t agno,
2179 xfs_agino_t next_agino,
2180 xfs_agino_t *old_next_agino)
2181{
2182 struct xfs_mount *mp = tp->t_mountp;
2183 struct xfs_dinode *dip;
2184 struct xfs_buf *ibp;
2185 xfs_agino_t old_value;
2186 int error;
2187
2188 ASSERT(xfs_verify_agino_or_null(mp, agno, next_agino));
2189
2190 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &dip, &ibp, 0, 0);
2191 if (error)
2192 return error;
2193
2194 /* Make sure the old pointer isn't garbage. */
2195 old_value = be32_to_cpu(dip->di_next_unlinked);
2196 if (!xfs_verify_agino_or_null(mp, agno, old_value)) {
2197 error = -EFSCORRUPTED;
2198 goto out;
2199 }
2200
2201 /*
2202 * Since we're updating a linked list, we should never find that the
2203 * current pointer is the same as the new value, unless we're
2204 * terminating the list.
2205 */
2206 *old_next_agino = old_value;
2207 if (old_value == next_agino) {
2208 if (next_agino != NULLAGINO)
2209 error = -EFSCORRUPTED;
2210 goto out;
2211 }
2212
2213 /* Ok, update the new pointer. */
2214 xfs_iunlink_update_dinode(tp, agno, XFS_INO_TO_AGINO(mp, ip->i_ino),
2215 ibp, dip, &ip->i_imap, next_agino);
2216 return 0;
2217out:
2218 xfs_trans_brelse(tp, ibp);
2219 return error;
2220}
2221
2222/*
2223 * This is called when the inode's link count has gone to 0 or we are creating
2224 * a tmpfile via O_TMPFILE. The inode @ip must have nlink == 0.
2225 *
2226 * We place the on-disk inode on a list in the AGI. It will be pulled from this
2227 * list when the inode is freed.
2228 */
2229STATIC int
2230xfs_iunlink(
2231 struct xfs_trans *tp,
2232 struct xfs_inode *ip)
2233{
2234 struct xfs_mount *mp = tp->t_mountp;
2235 struct xfs_agi *agi;
2236 struct xfs_buf *agibp;
2237 xfs_agino_t next_agino;
2238 xfs_agnumber_t agno = XFS_INO_TO_AGNO(mp, ip->i_ino);
2239 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2240 short bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
2241 int error;
2242
2243 ASSERT(VFS_I(ip)->i_nlink == 0);
2244 ASSERT(VFS_I(ip)->i_mode != 0);
2245 trace_xfs_iunlink(ip);
2246
2247 /* Get the agi buffer first. It ensures lock ordering on the list. */
2248 error = xfs_read_agi(mp, tp, agno, &agibp);
2249 if (error)
2250 return error;
2251 agi = XFS_BUF_TO_AGI(agibp);
2252
2253 /*
2254 * Get the index into the agi hash table for the list this inode will
2255 * go on. Make sure the pointer isn't garbage and that this inode
2256 * isn't already on the list.
2257 */
2258 next_agino = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2259 if (next_agino == agino ||
2260 !xfs_verify_agino_or_null(mp, agno, next_agino))
2261 return -EFSCORRUPTED;
2262
2263 if (next_agino != NULLAGINO) {
2264 struct xfs_perag *pag;
2265 xfs_agino_t old_agino;
2266
2267 /*
2268 * There is already another inode in the bucket, so point this
2269 * inode to the current head of the list.
2270 */
2271 error = xfs_iunlink_update_inode(tp, ip, agno, next_agino,
2272 &old_agino);
2273 if (error)
2274 return error;
2275 ASSERT(old_agino == NULLAGINO);
2276
2277 /*
2278 * agino has been unlinked, add a backref from the next inode
2279 * back to agino.
2280 */
2281 pag = xfs_perag_get(mp, agno);
2282 error = xfs_iunlink_add_backref(pag, agino, next_agino);
2283 xfs_perag_put(pag);
2284 if (error)
2285 return error;
2286 }
2287
2288 /* Point the head of the list to point to this inode. */
2289 return xfs_iunlink_update_bucket(tp, agno, agibp, bucket_index, agino);
2290}
2291
2292/* Return the imap, dinode pointer, and buffer for an inode. */
2293STATIC int
2294xfs_iunlink_map_ino(
2295 struct xfs_trans *tp,
2296 xfs_agnumber_t agno,
2297 xfs_agino_t agino,
2298 struct xfs_imap *imap,
2299 struct xfs_dinode **dipp,
2300 struct xfs_buf **bpp)
2301{
2302 struct xfs_mount *mp = tp->t_mountp;
2303 int error;
2304
2305 imap->im_blkno = 0;
2306 error = xfs_imap(mp, tp, XFS_AGINO_TO_INO(mp, agno, agino), imap, 0);
2307 if (error) {
2308 xfs_warn(mp, "%s: xfs_imap returned error %d.",
2309 __func__, error);
2310 return error;
2311 }
2312
2313 error = xfs_imap_to_bp(mp, tp, imap, dipp, bpp, 0, 0);
2314 if (error) {
2315 xfs_warn(mp, "%s: xfs_imap_to_bp returned error %d.",
2316 __func__, error);
2317 return error;
2318 }
2319
2320 return 0;
2321}
2322
2323/*
2324 * Walk the unlinked chain from @head_agino until we find the inode that
2325 * points to @target_agino. Return the inode number, map, dinode pointer,
2326 * and inode cluster buffer of that inode as @agino, @imap, @dipp, and @bpp.
2327 *
2328 * @tp, @pag, @head_agino, and @target_agino are input parameters.
2329 * @agino, @imap, @dipp, and @bpp are all output parameters.
2330 *
2331 * Do not call this function if @target_agino is the head of the list.
2332 */
2333STATIC int
2334xfs_iunlink_map_prev(
2335 struct xfs_trans *tp,
2336 xfs_agnumber_t agno,
2337 xfs_agino_t head_agino,
2338 xfs_agino_t target_agino,
2339 xfs_agino_t *agino,
2340 struct xfs_imap *imap,
2341 struct xfs_dinode **dipp,
2342 struct xfs_buf **bpp,
2343 struct xfs_perag *pag)
2344{
2345 struct xfs_mount *mp = tp->t_mountp;
2346 xfs_agino_t next_agino;
2347 int error;
2348
2349 ASSERT(head_agino != target_agino);
2350 *bpp = NULL;
2351
2352 /* See if our backref cache can find it faster. */
2353 *agino = xfs_iunlink_lookup_backref(pag, target_agino);
2354 if (*agino != NULLAGINO) {
2355 error = xfs_iunlink_map_ino(tp, agno, *agino, imap, dipp, bpp);
2356 if (error)
2357 return error;
2358
2359 if (be32_to_cpu((*dipp)->di_next_unlinked) == target_agino)
2360 return 0;
2361
2362 /*
2363 * If we get here the cache contents were corrupt, so drop the
2364 * buffer and fall back to walking the bucket list.
2365 */
2366 xfs_trans_brelse(tp, *bpp);
2367 *bpp = NULL;
2368 WARN_ON_ONCE(1);
2369 }
2370
2371 trace_xfs_iunlink_map_prev_fallback(mp, agno);
2372
2373 /* Otherwise, walk the entire bucket until we find it. */
2374 next_agino = head_agino;
2375 while (next_agino != target_agino) {
2376 xfs_agino_t unlinked_agino;
2377
2378 if (*bpp)
2379 xfs_trans_brelse(tp, *bpp);
2380
2381 *agino = next_agino;
2382 error = xfs_iunlink_map_ino(tp, agno, next_agino, imap, dipp,
2383 bpp);
2384 if (error)
2385 return error;
2386
2387 unlinked_agino = be32_to_cpu((*dipp)->di_next_unlinked);
2388 /*
2389 * Make sure this pointer is valid and isn't an obvious
2390 * infinite loop.
2391 */
2392 if (!xfs_verify_agino(mp, agno, unlinked_agino) ||
2393 next_agino == unlinked_agino) {
2394 XFS_CORRUPTION_ERROR(__func__,
2395 XFS_ERRLEVEL_LOW, mp,
2396 *dipp, sizeof(**dipp));
2397 error = -EFSCORRUPTED;
2398 return error;
2399 }
2400 next_agino = unlinked_agino;
2401 }
2402
2403 return 0;
2404}
2405
2406/*
2407 * Pull the on-disk inode from the AGI unlinked list.
2408 */
2409STATIC int
2410xfs_iunlink_remove(
2411 struct xfs_trans *tp,
2412 struct xfs_inode *ip)
2413{
2414 struct xfs_mount *mp = tp->t_mountp;
2415 struct xfs_agi *agi;
2416 struct xfs_buf *agibp;
2417 struct xfs_buf *last_ibp;
2418 struct xfs_dinode *last_dip = NULL;
2419 struct xfs_perag *pag = NULL;
2420 xfs_agnumber_t agno = XFS_INO_TO_AGNO(mp, ip->i_ino);
2421 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2422 xfs_agino_t next_agino;
2423 xfs_agino_t head_agino;
2424 short bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
2425 int error;
2426
2427 trace_xfs_iunlink_remove(ip);
2428
2429 /* Get the agi buffer first. It ensures lock ordering on the list. */
2430 error = xfs_read_agi(mp, tp, agno, &agibp);
2431 if (error)
2432 return error;
2433 agi = XFS_BUF_TO_AGI(agibp);
2434
2435 /*
2436 * Get the index into the agi hash table for the list this inode will
2437 * go on. Make sure the head pointer isn't garbage.
2438 */
2439 head_agino = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2440 if (!xfs_verify_agino(mp, agno, head_agino)) {
2441 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
2442 agi, sizeof(*agi));
2443 return -EFSCORRUPTED;
2444 }
2445
2446 /*
2447 * Set our inode's next_unlinked pointer to NULL and then return
2448 * the old pointer value so that we can update whatever was previous
2449 * to us in the list to point to whatever was next in the list.
2450 */
2451 error = xfs_iunlink_update_inode(tp, ip, agno, NULLAGINO, &next_agino);
2452 if (error)
2453 return error;
2454
2455 /*
2456 * If there was a backref pointing from the next inode back to this
2457 * one, remove it because we've removed this inode from the list.
2458 *
2459 * Later, if this inode was in the middle of the list we'll update
2460 * this inode's backref to point from the next inode.
2461 */
2462 if (next_agino != NULLAGINO) {
2463 pag = xfs_perag_get(mp, agno);
2464 error = xfs_iunlink_change_backref(pag, next_agino,
2465 NULLAGINO);
2466 if (error)
2467 goto out;
2468 }
2469
2470 if (head_agino == agino) {
2471 /* Point the head of the list to the next unlinked inode. */
2472 error = xfs_iunlink_update_bucket(tp, agno, agibp, bucket_index,
2473 next_agino);
2474 if (error)
2475 goto out;
2476 } else {
2477 struct xfs_imap imap;
2478 xfs_agino_t prev_agino;
2479
2480 if (!pag)
2481 pag = xfs_perag_get(mp, agno);
2482
2483 /* We need to search the list for the inode being freed. */
2484 error = xfs_iunlink_map_prev(tp, agno, head_agino, agino,
2485 &prev_agino, &imap, &last_dip, &last_ibp,
2486 pag);
2487 if (error)
2488 goto out;
2489
2490 /* Point the previous inode on the list to the next inode. */
2491 xfs_iunlink_update_dinode(tp, agno, prev_agino, last_ibp,
2492 last_dip, &imap, next_agino);
2493
2494 /*
2495 * Now we deal with the backref for this inode. If this inode
2496 * pointed at a real inode, change the backref that pointed to
2497 * us to point to our old next. If this inode was the end of
2498 * the list, delete the backref that pointed to us. Note that
2499 * change_backref takes care of deleting the backref if
2500 * next_agino is NULLAGINO.
2501 */
2502 error = xfs_iunlink_change_backref(pag, agino, next_agino);
2503 if (error)
2504 goto out;
2505 }
2506
2507out:
2508 if (pag)
2509 xfs_perag_put(pag);
2510 return error;
2511}
2512
2513/*
2514 * A big issue when freeing the inode cluster is that we _cannot_ skip any
2515 * inodes that are in memory - they all must be marked stale and attached to
2516 * the cluster buffer.
2517 */
2518STATIC int
2519xfs_ifree_cluster(
2520 xfs_inode_t *free_ip,
2521 xfs_trans_t *tp,
2522 struct xfs_icluster *xic)
2523{
2524 xfs_mount_t *mp = free_ip->i_mount;
2525 int nbufs;
2526 int i, j;
2527 int ioffset;
2528 xfs_daddr_t blkno;
2529 xfs_buf_t *bp;
2530 xfs_inode_t *ip;
2531 xfs_inode_log_item_t *iip;
2532 struct xfs_log_item *lip;
2533 struct xfs_perag *pag;
2534 struct xfs_ino_geometry *igeo = M_IGEO(mp);
2535 xfs_ino_t inum;
2536
2537 inum = xic->first_ino;
2538 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, inum));
2539 nbufs = igeo->ialloc_blks / igeo->blocks_per_cluster;
2540
2541 for (j = 0; j < nbufs; j++, inum += igeo->inodes_per_cluster) {
2542 /*
2543 * The allocation bitmap tells us which inodes of the chunk were
2544 * physically allocated. Skip the cluster if an inode falls into
2545 * a sparse region.
2546 */
2547 ioffset = inum - xic->first_ino;
2548 if ((xic->alloc & XFS_INOBT_MASK(ioffset)) == 0) {
2549 ASSERT(ioffset % igeo->inodes_per_cluster == 0);
2550 continue;
2551 }
2552
2553 blkno = XFS_AGB_TO_DADDR(mp, XFS_INO_TO_AGNO(mp, inum),
2554 XFS_INO_TO_AGBNO(mp, inum));
2555
2556 /*
2557 * We obtain and lock the backing buffer first in the process
2558 * here, as we have to ensure that any dirty inode that we
2559 * can't get the flush lock on is attached to the buffer.
2560 * If we scan the in-memory inodes first, then buffer IO can
2561 * complete before we get a lock on it, and hence we may fail
2562 * to mark all the active inodes on the buffer stale.
2563 */
2564 bp = xfs_trans_get_buf(tp, mp->m_ddev_targp, blkno,
2565 mp->m_bsize * igeo->blocks_per_cluster,
2566 XBF_UNMAPPED);
2567
2568 if (!bp)
2569 return -ENOMEM;
2570
2571 /*
2572 * This buffer may not have been correctly initialised as we
2573 * didn't read it from disk. That's not important because we are
2574 * only using to mark the buffer as stale in the log, and to
2575 * attach stale cached inodes on it. That means it will never be
2576 * dispatched for IO. If it is, we want to know about it, and we
2577 * want it to fail. We can acheive this by adding a write
2578 * verifier to the buffer.
2579 */
2580 bp->b_ops = &xfs_inode_buf_ops;
2581
2582 /*
2583 * Walk the inodes already attached to the buffer and mark them
2584 * stale. These will all have the flush locks held, so an
2585 * in-memory inode walk can't lock them. By marking them all
2586 * stale first, we will not attempt to lock them in the loop
2587 * below as the XFS_ISTALE flag will be set.
2588 */
2589 list_for_each_entry(lip, &bp->b_li_list, li_bio_list) {
2590 if (lip->li_type == XFS_LI_INODE) {
2591 iip = (xfs_inode_log_item_t *)lip;
2592 ASSERT(iip->ili_logged == 1);
2593 lip->li_cb = xfs_istale_done;
2594 xfs_trans_ail_copy_lsn(mp->m_ail,
2595 &iip->ili_flush_lsn,
2596 &iip->ili_item.li_lsn);
2597 xfs_iflags_set(iip->ili_inode, XFS_ISTALE);
2598 }
2599 }
2600
2601
2602 /*
2603 * For each inode in memory attempt to add it to the inode
2604 * buffer and set it up for being staled on buffer IO
2605 * completion. This is safe as we've locked out tail pushing
2606 * and flushing by locking the buffer.
2607 *
2608 * We have already marked every inode that was part of a
2609 * transaction stale above, which means there is no point in
2610 * even trying to lock them.
2611 */
2612 for (i = 0; i < igeo->inodes_per_cluster; i++) {
2613retry:
2614 rcu_read_lock();
2615 ip = radix_tree_lookup(&pag->pag_ici_root,
2616 XFS_INO_TO_AGINO(mp, (inum + i)));
2617
2618 /* Inode not in memory, nothing to do */
2619 if (!ip) {
2620 rcu_read_unlock();
2621 continue;
2622 }
2623
2624 /*
2625 * because this is an RCU protected lookup, we could
2626 * find a recently freed or even reallocated inode
2627 * during the lookup. We need to check under the
2628 * i_flags_lock for a valid inode here. Skip it if it
2629 * is not valid, the wrong inode or stale.
2630 */
2631 spin_lock(&ip->i_flags_lock);
2632 if (ip->i_ino != inum + i ||
2633 __xfs_iflags_test(ip, XFS_ISTALE)) {
2634 spin_unlock(&ip->i_flags_lock);
2635 rcu_read_unlock();
2636 continue;
2637 }
2638 spin_unlock(&ip->i_flags_lock);
2639
2640 /*
2641 * Don't try to lock/unlock the current inode, but we
2642 * _cannot_ skip the other inodes that we did not find
2643 * in the list attached to the buffer and are not
2644 * already marked stale. If we can't lock it, back off
2645 * and retry.
2646 */
2647 if (ip != free_ip) {
2648 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL)) {
2649 rcu_read_unlock();
2650 delay(1);
2651 goto retry;
2652 }
2653
2654 /*
2655 * Check the inode number again in case we're
2656 * racing with freeing in xfs_reclaim_inode().
2657 * See the comments in that function for more
2658 * information as to why the initial check is
2659 * not sufficient.
2660 */
2661 if (ip->i_ino != inum + i) {
2662 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2663 rcu_read_unlock();
2664 continue;
2665 }
2666 }
2667 rcu_read_unlock();
2668
2669 xfs_iflock(ip);
2670 xfs_iflags_set(ip, XFS_ISTALE);
2671
2672 /*
2673 * we don't need to attach clean inodes or those only
2674 * with unlogged changes (which we throw away, anyway).
2675 */
2676 iip = ip->i_itemp;
2677 if (!iip || xfs_inode_clean(ip)) {
2678 ASSERT(ip != free_ip);
2679 xfs_ifunlock(ip);
2680 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2681 continue;
2682 }
2683
2684 iip->ili_last_fields = iip->ili_fields;
2685 iip->ili_fields = 0;
2686 iip->ili_fsync_fields = 0;
2687 iip->ili_logged = 1;
2688 xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
2689 &iip->ili_item.li_lsn);
2690
2691 xfs_buf_attach_iodone(bp, xfs_istale_done,
2692 &iip->ili_item);
2693
2694 if (ip != free_ip)
2695 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2696 }
2697
2698 xfs_trans_stale_inode_buf(tp, bp);
2699 xfs_trans_binval(tp, bp);
2700 }
2701
2702 xfs_perag_put(pag);
2703 return 0;
2704}
2705
2706/*
2707 * Free any local-format buffers sitting around before we reset to
2708 * extents format.
2709 */
2710static inline void
2711xfs_ifree_local_data(
2712 struct xfs_inode *ip,
2713 int whichfork)
2714{
2715 struct xfs_ifork *ifp;
2716
2717 if (XFS_IFORK_FORMAT(ip, whichfork) != XFS_DINODE_FMT_LOCAL)
2718 return;
2719
2720 ifp = XFS_IFORK_PTR(ip, whichfork);
2721 xfs_idata_realloc(ip, -ifp->if_bytes, whichfork);
2722}
2723
2724/*
2725 * This is called to return an inode to the inode free list.
2726 * The inode should already be truncated to 0 length and have
2727 * no pages associated with it. This routine also assumes that
2728 * the inode is already a part of the transaction.
2729 *
2730 * The on-disk copy of the inode will have been added to the list
2731 * of unlinked inodes in the AGI. We need to remove the inode from
2732 * that list atomically with respect to freeing it here.
2733 */
2734int
2735xfs_ifree(
2736 struct xfs_trans *tp,
2737 struct xfs_inode *ip)
2738{
2739 int error;
2740 struct xfs_icluster xic = { 0 };
2741
2742 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
2743 ASSERT(VFS_I(ip)->i_nlink == 0);
2744 ASSERT(ip->i_d.di_nextents == 0);
2745 ASSERT(ip->i_d.di_anextents == 0);
2746 ASSERT(ip->i_d.di_size == 0 || !S_ISREG(VFS_I(ip)->i_mode));
2747 ASSERT(ip->i_d.di_nblocks == 0);
2748
2749 /*
2750 * Pull the on-disk inode from the AGI unlinked list.
2751 */
2752 error = xfs_iunlink_remove(tp, ip);
2753 if (error)
2754 return error;
2755
2756 error = xfs_difree(tp, ip->i_ino, &xic);
2757 if (error)
2758 return error;
2759
2760 xfs_ifree_local_data(ip, XFS_DATA_FORK);
2761 xfs_ifree_local_data(ip, XFS_ATTR_FORK);
2762
2763 VFS_I(ip)->i_mode = 0; /* mark incore inode as free */
2764 ip->i_d.di_flags = 0;
2765 ip->i_d.di_flags2 = 0;
2766 ip->i_d.di_dmevmask = 0;
2767 ip->i_d.di_forkoff = 0; /* mark the attr fork not in use */
2768 ip->i_d.di_format = XFS_DINODE_FMT_EXTENTS;
2769 ip->i_d.di_aformat = XFS_DINODE_FMT_EXTENTS;
2770
2771 /* Don't attempt to replay owner changes for a deleted inode */
2772 ip->i_itemp->ili_fields &= ~(XFS_ILOG_AOWNER|XFS_ILOG_DOWNER);
2773
2774 /*
2775 * Bump the generation count so no one will be confused
2776 * by reincarnations of this inode.
2777 */
2778 VFS_I(ip)->i_generation++;
2779 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2780
2781 if (xic.deleted)
2782 error = xfs_ifree_cluster(ip, tp, &xic);
2783
2784 return error;
2785}
2786
2787/*
2788 * This is called to unpin an inode. The caller must have the inode locked
2789 * in at least shared mode so that the buffer cannot be subsequently pinned
2790 * once someone is waiting for it to be unpinned.
2791 */
2792static void
2793xfs_iunpin(
2794 struct xfs_inode *ip)
2795{
2796 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
2797
2798 trace_xfs_inode_unpin_nowait(ip, _RET_IP_);
2799
2800 /* Give the log a push to start the unpinning I/O */
2801 xfs_log_force_lsn(ip->i_mount, ip->i_itemp->ili_last_lsn, 0, NULL);
2802
2803}
2804
2805static void
2806__xfs_iunpin_wait(
2807 struct xfs_inode *ip)
2808{
2809 wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_IPINNED_BIT);
2810 DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IPINNED_BIT);
2811
2812 xfs_iunpin(ip);
2813
2814 do {
2815 prepare_to_wait(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
2816 if (xfs_ipincount(ip))
2817 io_schedule();
2818 } while (xfs_ipincount(ip));
2819 finish_wait(wq, &wait.wq_entry);
2820}
2821
2822void
2823xfs_iunpin_wait(
2824 struct xfs_inode *ip)
2825{
2826 if (xfs_ipincount(ip))
2827 __xfs_iunpin_wait(ip);
2828}
2829
2830/*
2831 * Removing an inode from the namespace involves removing the directory entry
2832 * and dropping the link count on the inode. Removing the directory entry can
2833 * result in locking an AGF (directory blocks were freed) and removing a link
2834 * count can result in placing the inode on an unlinked list which results in
2835 * locking an AGI.
2836 *
2837 * The big problem here is that we have an ordering constraint on AGF and AGI
2838 * locking - inode allocation locks the AGI, then can allocate a new extent for
2839 * new inodes, locking the AGF after the AGI. Similarly, freeing the inode
2840 * removes the inode from the unlinked list, requiring that we lock the AGI
2841 * first, and then freeing the inode can result in an inode chunk being freed
2842 * and hence freeing disk space requiring that we lock an AGF.
2843 *
2844 * Hence the ordering that is imposed by other parts of the code is AGI before
2845 * AGF. This means we cannot remove the directory entry before we drop the inode
2846 * reference count and put it on the unlinked list as this results in a lock
2847 * order of AGF then AGI, and this can deadlock against inode allocation and
2848 * freeing. Therefore we must drop the link counts before we remove the
2849 * directory entry.
2850 *
2851 * This is still safe from a transactional point of view - it is not until we
2852 * get to xfs_defer_finish() that we have the possibility of multiple
2853 * transactions in this operation. Hence as long as we remove the directory
2854 * entry and drop the link count in the first transaction of the remove
2855 * operation, there are no transactional constraints on the ordering here.
2856 */
2857int
2858xfs_remove(
2859 xfs_inode_t *dp,
2860 struct xfs_name *name,
2861 xfs_inode_t *ip)
2862{
2863 xfs_mount_t *mp = dp->i_mount;
2864 xfs_trans_t *tp = NULL;
2865 int is_dir = S_ISDIR(VFS_I(ip)->i_mode);
2866 int error = 0;
2867 uint resblks;
2868
2869 trace_xfs_remove(dp, name);
2870
2871 if (XFS_FORCED_SHUTDOWN(mp))
2872 return -EIO;
2873
2874 error = xfs_qm_dqattach(dp);
2875 if (error)
2876 goto std_return;
2877
2878 error = xfs_qm_dqattach(ip);
2879 if (error)
2880 goto std_return;
2881
2882 /*
2883 * We try to get the real space reservation first,
2884 * allowing for directory btree deletion(s) implying
2885 * possible bmap insert(s). If we can't get the space
2886 * reservation then we use 0 instead, and avoid the bmap
2887 * btree insert(s) in the directory code by, if the bmap
2888 * insert tries to happen, instead trimming the LAST
2889 * block from the directory.
2890 */
2891 resblks = XFS_REMOVE_SPACE_RES(mp);
2892 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_remove, resblks, 0, 0, &tp);
2893 if (error == -ENOSPC) {
2894 resblks = 0;
2895 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_remove, 0, 0, 0,
2896 &tp);
2897 }
2898 if (error) {
2899 ASSERT(error != -ENOSPC);
2900 goto std_return;
2901 }
2902
2903 xfs_lock_two_inodes(dp, XFS_ILOCK_EXCL, ip, XFS_ILOCK_EXCL);
2904
2905 xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
2906 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
2907
2908 /*
2909 * If we're removing a directory perform some additional validation.
2910 */
2911 if (is_dir) {
2912 ASSERT(VFS_I(ip)->i_nlink >= 2);
2913 if (VFS_I(ip)->i_nlink != 2) {
2914 error = -ENOTEMPTY;
2915 goto out_trans_cancel;
2916 }
2917 if (!xfs_dir_isempty(ip)) {
2918 error = -ENOTEMPTY;
2919 goto out_trans_cancel;
2920 }
2921
2922 /* Drop the link from ip's "..". */
2923 error = xfs_droplink(tp, dp);
2924 if (error)
2925 goto out_trans_cancel;
2926
2927 /* Drop the "." link from ip to self. */
2928 error = xfs_droplink(tp, ip);
2929 if (error)
2930 goto out_trans_cancel;
2931 } else {
2932 /*
2933 * When removing a non-directory we need to log the parent
2934 * inode here. For a directory this is done implicitly
2935 * by the xfs_droplink call for the ".." entry.
2936 */
2937 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
2938 }
2939 xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2940
2941 /* Drop the link from dp to ip. */
2942 error = xfs_droplink(tp, ip);
2943 if (error)
2944 goto out_trans_cancel;
2945
2946 error = xfs_dir_removename(tp, dp, name, ip->i_ino, resblks);
2947 if (error) {
2948 ASSERT(error != -ENOENT);
2949 goto out_trans_cancel;
2950 }
2951
2952 /*
2953 * If this is a synchronous mount, make sure that the
2954 * remove transaction goes to disk before returning to
2955 * the user.
2956 */
2957 if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
2958 xfs_trans_set_sync(tp);
2959
2960 error = xfs_trans_commit(tp);
2961 if (error)
2962 goto std_return;
2963
2964 if (is_dir && xfs_inode_is_filestream(ip))
2965 xfs_filestream_deassociate(ip);
2966
2967 return 0;
2968
2969 out_trans_cancel:
2970 xfs_trans_cancel(tp);
2971 std_return:
2972 return error;
2973}
2974
2975/*
2976 * Enter all inodes for a rename transaction into a sorted array.
2977 */
2978#define __XFS_SORT_INODES 5
2979STATIC void
2980xfs_sort_for_rename(
2981 struct xfs_inode *dp1, /* in: old (source) directory inode */
2982 struct xfs_inode *dp2, /* in: new (target) directory inode */
2983 struct xfs_inode *ip1, /* in: inode of old entry */
2984 struct xfs_inode *ip2, /* in: inode of new entry */
2985 struct xfs_inode *wip, /* in: whiteout inode */
2986 struct xfs_inode **i_tab,/* out: sorted array of inodes */
2987 int *num_inodes) /* in/out: inodes in array */
2988{
2989 int i, j;
2990
2991 ASSERT(*num_inodes == __XFS_SORT_INODES);
2992 memset(i_tab, 0, *num_inodes * sizeof(struct xfs_inode *));
2993
2994 /*
2995 * i_tab contains a list of pointers to inodes. We initialize
2996 * the table here & we'll sort it. We will then use it to
2997 * order the acquisition of the inode locks.
2998 *
2999 * Note that the table may contain duplicates. e.g., dp1 == dp2.
3000 */
3001 i = 0;
3002 i_tab[i++] = dp1;
3003 i_tab[i++] = dp2;
3004 i_tab[i++] = ip1;
3005 if (ip2)
3006 i_tab[i++] = ip2;
3007 if (wip)
3008 i_tab[i++] = wip;
3009 *num_inodes = i;
3010
3011 /*
3012 * Sort the elements via bubble sort. (Remember, there are at
3013 * most 5 elements to sort, so this is adequate.)
3014 */
3015 for (i = 0; i < *num_inodes; i++) {
3016 for (j = 1; j < *num_inodes; j++) {
3017 if (i_tab[j]->i_ino < i_tab[j-1]->i_ino) {
3018 struct xfs_inode *temp = i_tab[j];
3019 i_tab[j] = i_tab[j-1];
3020 i_tab[j-1] = temp;
3021 }
3022 }
3023 }
3024}
3025
3026static int
3027xfs_finish_rename(
3028 struct xfs_trans *tp)
3029{
3030 /*
3031 * If this is a synchronous mount, make sure that the rename transaction
3032 * goes to disk before returning to the user.
3033 */
3034 if (tp->t_mountp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
3035 xfs_trans_set_sync(tp);
3036
3037 return xfs_trans_commit(tp);
3038}
3039
3040/*
3041 * xfs_cross_rename()
3042 *
3043 * responsible for handling RENAME_EXCHANGE flag in renameat2() sytemcall
3044 */
3045STATIC int
3046xfs_cross_rename(
3047 struct xfs_trans *tp,
3048 struct xfs_inode *dp1,
3049 struct xfs_name *name1,
3050 struct xfs_inode *ip1,
3051 struct xfs_inode *dp2,
3052 struct xfs_name *name2,
3053 struct xfs_inode *ip2,
3054 int spaceres)
3055{
3056 int error = 0;
3057 int ip1_flags = 0;
3058 int ip2_flags = 0;
3059 int dp2_flags = 0;
3060
3061 /* Swap inode number for dirent in first parent */
3062 error = xfs_dir_replace(tp, dp1, name1, ip2->i_ino, spaceres);
3063 if (error)
3064 goto out_trans_abort;
3065
3066 /* Swap inode number for dirent in second parent */
3067 error = xfs_dir_replace(tp, dp2, name2, ip1->i_ino, spaceres);
3068 if (error)
3069 goto out_trans_abort;
3070
3071 /*
3072 * If we're renaming one or more directories across different parents,
3073 * update the respective ".." entries (and link counts) to match the new
3074 * parents.
3075 */
3076 if (dp1 != dp2) {
3077 dp2_flags = XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
3078
3079 if (S_ISDIR(VFS_I(ip2)->i_mode)) {
3080 error = xfs_dir_replace(tp, ip2, &xfs_name_dotdot,
3081 dp1->i_ino, spaceres);
3082 if (error)
3083 goto out_trans_abort;
3084
3085 /* transfer ip2 ".." reference to dp1 */
3086 if (!S_ISDIR(VFS_I(ip1)->i_mode)) {
3087 error = xfs_droplink(tp, dp2);
3088 if (error)
3089 goto out_trans_abort;
3090 xfs_bumplink(tp, dp1);
3091 }
3092
3093 /*
3094 * Although ip1 isn't changed here, userspace needs
3095 * to be warned about the change, so that applications
3096 * relying on it (like backup ones), will properly
3097 * notify the change
3098 */
3099 ip1_flags |= XFS_ICHGTIME_CHG;
3100 ip2_flags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
3101 }
3102
3103 if (S_ISDIR(VFS_I(ip1)->i_mode)) {
3104 error = xfs_dir_replace(tp, ip1, &xfs_name_dotdot,
3105 dp2->i_ino, spaceres);
3106 if (error)
3107 goto out_trans_abort;
3108
3109 /* transfer ip1 ".." reference to dp2 */
3110 if (!S_ISDIR(VFS_I(ip2)->i_mode)) {
3111 error = xfs_droplink(tp, dp1);
3112 if (error)
3113 goto out_trans_abort;
3114 xfs_bumplink(tp, dp2);
3115 }
3116
3117 /*
3118 * Although ip2 isn't changed here, userspace needs
3119 * to be warned about the change, so that applications
3120 * relying on it (like backup ones), will properly
3121 * notify the change
3122 */
3123 ip1_flags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
3124 ip2_flags |= XFS_ICHGTIME_CHG;
3125 }
3126 }
3127
3128 if (ip1_flags) {
3129 xfs_trans_ichgtime(tp, ip1, ip1_flags);
3130 xfs_trans_log_inode(tp, ip1, XFS_ILOG_CORE);
3131 }
3132 if (ip2_flags) {
3133 xfs_trans_ichgtime(tp, ip2, ip2_flags);
3134 xfs_trans_log_inode(tp, ip2, XFS_ILOG_CORE);
3135 }
3136 if (dp2_flags) {
3137 xfs_trans_ichgtime(tp, dp2, dp2_flags);
3138 xfs_trans_log_inode(tp, dp2, XFS_ILOG_CORE);
3139 }
3140 xfs_trans_ichgtime(tp, dp1, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3141 xfs_trans_log_inode(tp, dp1, XFS_ILOG_CORE);
3142 return xfs_finish_rename(tp);
3143
3144out_trans_abort:
3145 xfs_trans_cancel(tp);
3146 return error;
3147}
3148
3149/*
3150 * xfs_rename_alloc_whiteout()
3151 *
3152 * Return a referenced, unlinked, unlocked inode that that can be used as a
3153 * whiteout in a rename transaction. We use a tmpfile inode here so that if we
3154 * crash between allocating the inode and linking it into the rename transaction
3155 * recovery will free the inode and we won't leak it.
3156 */
3157static int
3158xfs_rename_alloc_whiteout(
3159 struct xfs_inode *dp,
3160 struct xfs_inode **wip)
3161{
3162 struct xfs_inode *tmpfile;
3163 int error;
3164
3165 error = xfs_create_tmpfile(dp, S_IFCHR | WHITEOUT_MODE, &tmpfile);
3166 if (error)
3167 return error;
3168
3169 /*
3170 * Prepare the tmpfile inode as if it were created through the VFS.
3171 * Complete the inode setup and flag it as linkable. nlink is already
3172 * zero, so we can skip the drop_nlink.
3173 */
3174 xfs_setup_iops(tmpfile);
3175 xfs_finish_inode_setup(tmpfile);
3176 VFS_I(tmpfile)->i_state |= I_LINKABLE;
3177
3178 *wip = tmpfile;
3179 return 0;
3180}
3181
3182/*
3183 * xfs_rename
3184 */
3185int
3186xfs_rename(
3187 struct xfs_inode *src_dp,
3188 struct xfs_name *src_name,
3189 struct xfs_inode *src_ip,
3190 struct xfs_inode *target_dp,
3191 struct xfs_name *target_name,
3192 struct xfs_inode *target_ip,
3193 unsigned int flags)
3194{
3195 struct xfs_mount *mp = src_dp->i_mount;
3196 struct xfs_trans *tp;
3197 struct xfs_inode *wip = NULL; /* whiteout inode */
3198 struct xfs_inode *inodes[__XFS_SORT_INODES];
3199 int num_inodes = __XFS_SORT_INODES;
3200 bool new_parent = (src_dp != target_dp);
3201 bool src_is_directory = S_ISDIR(VFS_I(src_ip)->i_mode);
3202 int spaceres;
3203 int error;
3204
3205 trace_xfs_rename(src_dp, target_dp, src_name, target_name);
3206
3207 if ((flags & RENAME_EXCHANGE) && !target_ip)
3208 return -EINVAL;
3209
3210 /*
3211 * If we are doing a whiteout operation, allocate the whiteout inode
3212 * we will be placing at the target and ensure the type is set
3213 * appropriately.
3214 */
3215 if (flags & RENAME_WHITEOUT) {
3216 ASSERT(!(flags & (RENAME_NOREPLACE | RENAME_EXCHANGE)));
3217 error = xfs_rename_alloc_whiteout(target_dp, &wip);
3218 if (error)
3219 return error;
3220
3221 /* setup target dirent info as whiteout */
3222 src_name->type = XFS_DIR3_FT_CHRDEV;
3223 }
3224
3225 xfs_sort_for_rename(src_dp, target_dp, src_ip, target_ip, wip,
3226 inodes, &num_inodes);
3227
3228 spaceres = XFS_RENAME_SPACE_RES(mp, target_name->len);
3229 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, spaceres, 0, 0, &tp);
3230 if (error == -ENOSPC) {
3231 spaceres = 0;
3232 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, 0, 0, 0,
3233 &tp);
3234 }
3235 if (error)
3236 goto out_release_wip;
3237
3238 /*
3239 * Attach the dquots to the inodes
3240 */
3241 error = xfs_qm_vop_rename_dqattach(inodes);
3242 if (error)
3243 goto out_trans_cancel;
3244
3245 /*
3246 * Lock all the participating inodes. Depending upon whether
3247 * the target_name exists in the target directory, and
3248 * whether the target directory is the same as the source
3249 * directory, we can lock from 2 to 4 inodes.
3250 */
3251 xfs_lock_inodes(inodes, num_inodes, XFS_ILOCK_EXCL);
3252
3253 /*
3254 * Join all the inodes to the transaction. From this point on,
3255 * we can rely on either trans_commit or trans_cancel to unlock
3256 * them.
3257 */
3258 xfs_trans_ijoin(tp, src_dp, XFS_ILOCK_EXCL);
3259 if (new_parent)
3260 xfs_trans_ijoin(tp, target_dp, XFS_ILOCK_EXCL);
3261 xfs_trans_ijoin(tp, src_ip, XFS_ILOCK_EXCL);
3262 if (target_ip)
3263 xfs_trans_ijoin(tp, target_ip, XFS_ILOCK_EXCL);
3264 if (wip)
3265 xfs_trans_ijoin(tp, wip, XFS_ILOCK_EXCL);
3266
3267 /*
3268 * If we are using project inheritance, we only allow renames
3269 * into our tree when the project IDs are the same; else the
3270 * tree quota mechanism would be circumvented.
3271 */
3272 if (unlikely((target_dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
3273 (xfs_get_projid(target_dp) != xfs_get_projid(src_ip)))) {
3274 error = -EXDEV;
3275 goto out_trans_cancel;
3276 }
3277
3278 /* RENAME_EXCHANGE is unique from here on. */
3279 if (flags & RENAME_EXCHANGE)
3280 return xfs_cross_rename(tp, src_dp, src_name, src_ip,
3281 target_dp, target_name, target_ip,
3282 spaceres);
3283
3284 /*
3285 * Check for expected errors before we dirty the transaction
3286 * so we can return an error without a transaction abort.
3287 */
3288 if (target_ip == NULL) {
3289 /*
3290 * If there's no space reservation, check the entry will
3291 * fit before actually inserting it.
3292 */
3293 if (!spaceres) {
3294 error = xfs_dir_canenter(tp, target_dp, target_name);
3295 if (error)
3296 goto out_trans_cancel;
3297 }
3298 } else {
3299 /*
3300 * If target exists and it's a directory, check that whether
3301 * it can be destroyed.
3302 */
3303 if (S_ISDIR(VFS_I(target_ip)->i_mode) &&
3304 (!xfs_dir_isempty(target_ip) ||
3305 (VFS_I(target_ip)->i_nlink > 2))) {
3306 error = -EEXIST;
3307 goto out_trans_cancel;
3308 }
3309 }
3310
3311 /*
3312 * Directory entry creation below may acquire the AGF. Remove
3313 * the whiteout from the unlinked list first to preserve correct
3314 * AGI/AGF locking order. This dirties the transaction so failures
3315 * after this point will abort and log recovery will clean up the
3316 * mess.
3317 *
3318 * For whiteouts, we need to bump the link count on the whiteout
3319 * inode. After this point, we have a real link, clear the tmpfile
3320 * state flag from the inode so it doesn't accidentally get misused
3321 * in future.
3322 */
3323 if (wip) {
3324 ASSERT(VFS_I(wip)->i_nlink == 0);
3325 error = xfs_iunlink_remove(tp, wip);
3326 if (error)
3327 goto out_trans_cancel;
3328
3329 xfs_bumplink(tp, wip);
3330 xfs_trans_log_inode(tp, wip, XFS_ILOG_CORE);
3331 VFS_I(wip)->i_state &= ~I_LINKABLE;
3332 }
3333
3334 /*
3335 * Set up the target.
3336 */
3337 if (target_ip == NULL) {
3338 /*
3339 * If target does not exist and the rename crosses
3340 * directories, adjust the target directory link count
3341 * to account for the ".." reference from the new entry.
3342 */
3343 error = xfs_dir_createname(tp, target_dp, target_name,
3344 src_ip->i_ino, spaceres);
3345 if (error)
3346 goto out_trans_cancel;
3347
3348 xfs_trans_ichgtime(tp, target_dp,
3349 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3350
3351 if (new_parent && src_is_directory) {
3352 xfs_bumplink(tp, target_dp);
3353 }
3354 } else { /* target_ip != NULL */
3355 /*
3356 * Link the source inode under the target name.
3357 * If the source inode is a directory and we are moving
3358 * it across directories, its ".." entry will be
3359 * inconsistent until we replace that down below.
3360 *
3361 * In case there is already an entry with the same
3362 * name at the destination directory, remove it first.
3363 */
3364 error = xfs_dir_replace(tp, target_dp, target_name,
3365 src_ip->i_ino, spaceres);
3366 if (error)
3367 goto out_trans_cancel;
3368
3369 xfs_trans_ichgtime(tp, target_dp,
3370 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3371
3372 /*
3373 * Decrement the link count on the target since the target
3374 * dir no longer points to it.
3375 */
3376 error = xfs_droplink(tp, target_ip);
3377 if (error)
3378 goto out_trans_cancel;
3379
3380 if (src_is_directory) {
3381 /*
3382 * Drop the link from the old "." entry.
3383 */
3384 error = xfs_droplink(tp, target_ip);
3385 if (error)
3386 goto out_trans_cancel;
3387 }
3388 } /* target_ip != NULL */
3389
3390 /*
3391 * Remove the source.
3392 */
3393 if (new_parent && src_is_directory) {
3394 /*
3395 * Rewrite the ".." entry to point to the new
3396 * directory.
3397 */
3398 error = xfs_dir_replace(tp, src_ip, &xfs_name_dotdot,
3399 target_dp->i_ino, spaceres);
3400 ASSERT(error != -EEXIST);
3401 if (error)
3402 goto out_trans_cancel;
3403 }
3404
3405 /*
3406 * We always want to hit the ctime on the source inode.
3407 *
3408 * This isn't strictly required by the standards since the source
3409 * inode isn't really being changed, but old unix file systems did
3410 * it and some incremental backup programs won't work without it.
3411 */
3412 xfs_trans_ichgtime(tp, src_ip, XFS_ICHGTIME_CHG);
3413 xfs_trans_log_inode(tp, src_ip, XFS_ILOG_CORE);
3414
3415 /*
3416 * Adjust the link count on src_dp. This is necessary when
3417 * renaming a directory, either within one parent when
3418 * the target existed, or across two parent directories.
3419 */
3420 if (src_is_directory && (new_parent || target_ip != NULL)) {
3421
3422 /*
3423 * Decrement link count on src_directory since the
3424 * entry that's moved no longer points to it.
3425 */
3426 error = xfs_droplink(tp, src_dp);
3427 if (error)
3428 goto out_trans_cancel;
3429 }
3430
3431 /*
3432 * For whiteouts, we only need to update the source dirent with the
3433 * inode number of the whiteout inode rather than removing it
3434 * altogether.
3435 */
3436 if (wip) {
3437 error = xfs_dir_replace(tp, src_dp, src_name, wip->i_ino,
3438 spaceres);
3439 } else
3440 error = xfs_dir_removename(tp, src_dp, src_name, src_ip->i_ino,
3441 spaceres);
3442 if (error)
3443 goto out_trans_cancel;
3444
3445 xfs_trans_ichgtime(tp, src_dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3446 xfs_trans_log_inode(tp, src_dp, XFS_ILOG_CORE);
3447 if (new_parent)
3448 xfs_trans_log_inode(tp, target_dp, XFS_ILOG_CORE);
3449
3450 error = xfs_finish_rename(tp);
3451 if (wip)
3452 xfs_irele(wip);
3453 return error;
3454
3455out_trans_cancel:
3456 xfs_trans_cancel(tp);
3457out_release_wip:
3458 if (wip)
3459 xfs_irele(wip);
3460 return error;
3461}
3462
3463STATIC int
3464xfs_iflush_cluster(
3465 struct xfs_inode *ip,
3466 struct xfs_buf *bp)
3467{
3468 struct xfs_mount *mp = ip->i_mount;
3469 struct xfs_perag *pag;
3470 unsigned long first_index, mask;
3471 int cilist_size;
3472 struct xfs_inode **cilist;
3473 struct xfs_inode *cip;
3474 struct xfs_ino_geometry *igeo = M_IGEO(mp);
3475 int nr_found;
3476 int clcount = 0;
3477 int i;
3478
3479 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
3480
3481 cilist_size = igeo->inodes_per_cluster * sizeof(struct xfs_inode *);
3482 cilist = kmem_alloc(cilist_size, KM_MAYFAIL|KM_NOFS);
3483 if (!cilist)
3484 goto out_put;
3485
3486 mask = ~(igeo->inodes_per_cluster - 1);
3487 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino) & mask;
3488 rcu_read_lock();
3489 /* really need a gang lookup range call here */
3490 nr_found = radix_tree_gang_lookup(&pag->pag_ici_root, (void**)cilist,
3491 first_index, igeo->inodes_per_cluster);
3492 if (nr_found == 0)
3493 goto out_free;
3494
3495 for (i = 0; i < nr_found; i++) {
3496 cip = cilist[i];
3497 if (cip == ip)
3498 continue;
3499
3500 /*
3501 * because this is an RCU protected lookup, we could find a
3502 * recently freed or even reallocated inode during the lookup.
3503 * We need to check under the i_flags_lock for a valid inode
3504 * here. Skip it if it is not valid or the wrong inode.
3505 */
3506 spin_lock(&cip->i_flags_lock);
3507 if (!cip->i_ino ||
3508 __xfs_iflags_test(cip, XFS_ISTALE)) {
3509 spin_unlock(&cip->i_flags_lock);
3510 continue;
3511 }
3512
3513 /*
3514 * Once we fall off the end of the cluster, no point checking
3515 * any more inodes in the list because they will also all be
3516 * outside the cluster.
3517 */
3518 if ((XFS_INO_TO_AGINO(mp, cip->i_ino) & mask) != first_index) {
3519 spin_unlock(&cip->i_flags_lock);
3520 break;
3521 }
3522 spin_unlock(&cip->i_flags_lock);
3523
3524 /*
3525 * Do an un-protected check to see if the inode is dirty and
3526 * is a candidate for flushing. These checks will be repeated
3527 * later after the appropriate locks are acquired.
3528 */
3529 if (xfs_inode_clean(cip) && xfs_ipincount(cip) == 0)
3530 continue;
3531
3532 /*
3533 * Try to get locks. If any are unavailable or it is pinned,
3534 * then this inode cannot be flushed and is skipped.
3535 */
3536
3537 if (!xfs_ilock_nowait(cip, XFS_ILOCK_SHARED))
3538 continue;
3539 if (!xfs_iflock_nowait(cip)) {
3540 xfs_iunlock(cip, XFS_ILOCK_SHARED);
3541 continue;
3542 }
3543 if (xfs_ipincount(cip)) {
3544 xfs_ifunlock(cip);
3545 xfs_iunlock(cip, XFS_ILOCK_SHARED);
3546 continue;
3547 }
3548
3549
3550 /*
3551 * Check the inode number again, just to be certain we are not
3552 * racing with freeing in xfs_reclaim_inode(). See the comments
3553 * in that function for more information as to why the initial
3554 * check is not sufficient.
3555 */
3556 if (!cip->i_ino) {
3557 xfs_ifunlock(cip);
3558 xfs_iunlock(cip, XFS_ILOCK_SHARED);
3559 continue;
3560 }
3561
3562 /*
3563 * arriving here means that this inode can be flushed. First
3564 * re-check that it's dirty before flushing.
3565 */
3566 if (!xfs_inode_clean(cip)) {
3567 int error;
3568 error = xfs_iflush_int(cip, bp);
3569 if (error) {
3570 xfs_iunlock(cip, XFS_ILOCK_SHARED);
3571 goto cluster_corrupt_out;
3572 }
3573 clcount++;
3574 } else {
3575 xfs_ifunlock(cip);
3576 }
3577 xfs_iunlock(cip, XFS_ILOCK_SHARED);
3578 }
3579
3580 if (clcount) {
3581 XFS_STATS_INC(mp, xs_icluster_flushcnt);
3582 XFS_STATS_ADD(mp, xs_icluster_flushinode, clcount);
3583 }
3584
3585out_free:
3586 rcu_read_unlock();
3587 kmem_free(cilist);
3588out_put:
3589 xfs_perag_put(pag);
3590 return 0;
3591
3592
3593cluster_corrupt_out:
3594 /*
3595 * Corruption detected in the clustering loop. Invalidate the
3596 * inode buffer and shut down the filesystem.
3597 */
3598 rcu_read_unlock();
3599
3600 /*
3601 * We'll always have an inode attached to the buffer for completion
3602 * process by the time we are called from xfs_iflush(). Hence we have
3603 * always need to do IO completion processing to abort the inodes
3604 * attached to the buffer. handle them just like the shutdown case in
3605 * xfs_buf_submit().
3606 */
3607 ASSERT(bp->b_iodone);
3608 bp->b_flags |= XBF_ASYNC;
3609 bp->b_flags &= ~XBF_DONE;
3610 xfs_buf_stale(bp);
3611 xfs_buf_ioerror(bp, -EIO);
3612 xfs_buf_ioend(bp);
3613
3614 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
3615
3616 /* abort the corrupt inode, as it was not attached to the buffer */
3617 xfs_iflush_abort(cip, false);
3618 kmem_free(cilist);
3619 xfs_perag_put(pag);
3620 return -EFSCORRUPTED;
3621}
3622
3623/*
3624 * Flush dirty inode metadata into the backing buffer.
3625 *
3626 * The caller must have the inode lock and the inode flush lock held. The
3627 * inode lock will still be held upon return to the caller, and the inode
3628 * flush lock will be released after the inode has reached the disk.
3629 *
3630 * The caller must write out the buffer returned in *bpp and release it.
3631 */
3632int
3633xfs_iflush(
3634 struct xfs_inode *ip,
3635 struct xfs_buf **bpp)
3636{
3637 struct xfs_mount *mp = ip->i_mount;
3638 struct xfs_buf *bp = NULL;
3639 struct xfs_dinode *dip;
3640 int error;
3641
3642 XFS_STATS_INC(mp, xs_iflush_count);
3643
3644 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
3645 ASSERT(xfs_isiflocked(ip));
3646 ASSERT(ip->i_d.di_format != XFS_DINODE_FMT_BTREE ||
3647 ip->i_d.di_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
3648
3649 *bpp = NULL;
3650
3651 xfs_iunpin_wait(ip);
3652
3653 /*
3654 * For stale inodes we cannot rely on the backing buffer remaining
3655 * stale in cache for the remaining life of the stale inode and so
3656 * xfs_imap_to_bp() below may give us a buffer that no longer contains
3657 * inodes below. We have to check this after ensuring the inode is
3658 * unpinned so that it is safe to reclaim the stale inode after the
3659 * flush call.
3660 */
3661 if (xfs_iflags_test(ip, XFS_ISTALE)) {
3662 xfs_ifunlock(ip);
3663 return 0;
3664 }
3665
3666 /*
3667 * This may have been unpinned because the filesystem is shutting
3668 * down forcibly. If that's the case we must not write this inode
3669 * to disk, because the log record didn't make it to disk.
3670 *
3671 * We also have to remove the log item from the AIL in this case,
3672 * as we wait for an empty AIL as part of the unmount process.
3673 */
3674 if (XFS_FORCED_SHUTDOWN(mp)) {
3675 error = -EIO;
3676 goto abort_out;
3677 }
3678
3679 /*
3680 * Get the buffer containing the on-disk inode. We are doing a try-lock
3681 * operation here, so we may get an EAGAIN error. In that case, we
3682 * simply want to return with the inode still dirty.
3683 *
3684 * If we get any other error, we effectively have a corruption situation
3685 * and we cannot flush the inode, so we treat it the same as failing
3686 * xfs_iflush_int().
3687 */
3688 error = xfs_imap_to_bp(mp, NULL, &ip->i_imap, &dip, &bp, XBF_TRYLOCK,
3689 0);
3690 if (error == -EAGAIN) {
3691 xfs_ifunlock(ip);
3692 return error;
3693 }
3694 if (error)
3695 goto corrupt_out;
3696
3697 /*
3698 * First flush out the inode that xfs_iflush was called with.
3699 */
3700 error = xfs_iflush_int(ip, bp);
3701 if (error)
3702 goto corrupt_out;
3703
3704 /*
3705 * If the buffer is pinned then push on the log now so we won't
3706 * get stuck waiting in the write for too long.
3707 */
3708 if (xfs_buf_ispinned(bp))
3709 xfs_log_force(mp, 0);
3710
3711 /*
3712 * inode clustering: try to gather other inodes into this write
3713 *
3714 * Note: Any error during clustering will result in the filesystem
3715 * being shut down and completion callbacks run on the cluster buffer.
3716 * As we have already flushed and attached this inode to the buffer,
3717 * it has already been aborted and released by xfs_iflush_cluster() and
3718 * so we have no further error handling to do here.
3719 */
3720 error = xfs_iflush_cluster(ip, bp);
3721 if (error)
3722 return error;
3723
3724 *bpp = bp;
3725 return 0;
3726
3727corrupt_out:
3728 if (bp)
3729 xfs_buf_relse(bp);
3730 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
3731abort_out:
3732 /* abort the corrupt inode, as it was not attached to the buffer */
3733 xfs_iflush_abort(ip, false);
3734 return error;
3735}
3736
3737/*
3738 * If there are inline format data / attr forks attached to this inode,
3739 * make sure they're not corrupt.
3740 */
3741bool
3742xfs_inode_verify_forks(
3743 struct xfs_inode *ip)
3744{
3745 struct xfs_ifork *ifp;
3746 xfs_failaddr_t fa;
3747
3748 fa = xfs_ifork_verify_data(ip, &xfs_default_ifork_ops);
3749 if (fa) {
3750 ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
3751 xfs_inode_verifier_error(ip, -EFSCORRUPTED, "data fork",
3752 ifp->if_u1.if_data, ifp->if_bytes, fa);
3753 return false;
3754 }
3755
3756 fa = xfs_ifork_verify_attr(ip, &xfs_default_ifork_ops);
3757 if (fa) {
3758 ifp = XFS_IFORK_PTR(ip, XFS_ATTR_FORK);
3759 xfs_inode_verifier_error(ip, -EFSCORRUPTED, "attr fork",
3760 ifp ? ifp->if_u1.if_data : NULL,
3761 ifp ? ifp->if_bytes : 0, fa);
3762 return false;
3763 }
3764 return true;
3765}
3766
3767STATIC int
3768xfs_iflush_int(
3769 struct xfs_inode *ip,
3770 struct xfs_buf *bp)
3771{
3772 struct xfs_inode_log_item *iip = ip->i_itemp;
3773 struct xfs_dinode *dip;
3774 struct xfs_mount *mp = ip->i_mount;
3775
3776 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
3777 ASSERT(xfs_isiflocked(ip));
3778 ASSERT(ip->i_d.di_format != XFS_DINODE_FMT_BTREE ||
3779 ip->i_d.di_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
3780 ASSERT(iip != NULL && iip->ili_fields != 0);
3781 ASSERT(ip->i_d.di_version > 1);
3782
3783 /* set *dip = inode's place in the buffer */
3784 dip = xfs_buf_offset(bp, ip->i_imap.im_boffset);
3785
3786 if (XFS_TEST_ERROR(dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC),
3787 mp, XFS_ERRTAG_IFLUSH_1)) {
3788 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3789 "%s: Bad inode %Lu magic number 0x%x, ptr "PTR_FMT,
3790 __func__, ip->i_ino, be16_to_cpu(dip->di_magic), dip);
3791 goto corrupt_out;
3792 }
3793 if (S_ISREG(VFS_I(ip)->i_mode)) {
3794 if (XFS_TEST_ERROR(
3795 (ip->i_d.di_format != XFS_DINODE_FMT_EXTENTS) &&
3796 (ip->i_d.di_format != XFS_DINODE_FMT_BTREE),
3797 mp, XFS_ERRTAG_IFLUSH_3)) {
3798 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3799 "%s: Bad regular inode %Lu, ptr "PTR_FMT,
3800 __func__, ip->i_ino, ip);
3801 goto corrupt_out;
3802 }
3803 } else if (S_ISDIR(VFS_I(ip)->i_mode)) {
3804 if (XFS_TEST_ERROR(
3805 (ip->i_d.di_format != XFS_DINODE_FMT_EXTENTS) &&
3806 (ip->i_d.di_format != XFS_DINODE_FMT_BTREE) &&
3807 (ip->i_d.di_format != XFS_DINODE_FMT_LOCAL),
3808 mp, XFS_ERRTAG_IFLUSH_4)) {
3809 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3810 "%s: Bad directory inode %Lu, ptr "PTR_FMT,
3811 __func__, ip->i_ino, ip);
3812 goto corrupt_out;
3813 }
3814 }
3815 if (XFS_TEST_ERROR(ip->i_d.di_nextents + ip->i_d.di_anextents >
3816 ip->i_d.di_nblocks, mp, XFS_ERRTAG_IFLUSH_5)) {
3817 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3818 "%s: detected corrupt incore inode %Lu, "
3819 "total extents = %d, nblocks = %Ld, ptr "PTR_FMT,
3820 __func__, ip->i_ino,
3821 ip->i_d.di_nextents + ip->i_d.di_anextents,
3822 ip->i_d.di_nblocks, ip);
3823 goto corrupt_out;
3824 }
3825 if (XFS_TEST_ERROR(ip->i_d.di_forkoff > mp->m_sb.sb_inodesize,
3826 mp, XFS_ERRTAG_IFLUSH_6)) {
3827 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3828 "%s: bad inode %Lu, forkoff 0x%x, ptr "PTR_FMT,
3829 __func__, ip->i_ino, ip->i_d.di_forkoff, ip);
3830 goto corrupt_out;
3831 }
3832
3833 /*
3834 * Inode item log recovery for v2 inodes are dependent on the
3835 * di_flushiter count for correct sequencing. We bump the flush
3836 * iteration count so we can detect flushes which postdate a log record
3837 * during recovery. This is redundant as we now log every change and
3838 * hence this can't happen but we need to still do it to ensure
3839 * backwards compatibility with old kernels that predate logging all
3840 * inode changes.
3841 */
3842 if (ip->i_d.di_version < 3)
3843 ip->i_d.di_flushiter++;
3844
3845 /* Check the inline fork data before we write out. */
3846 if (!xfs_inode_verify_forks(ip))
3847 goto corrupt_out;
3848
3849 /*
3850 * Copy the dirty parts of the inode into the on-disk inode. We always
3851 * copy out the core of the inode, because if the inode is dirty at all
3852 * the core must be.
3853 */
3854 xfs_inode_to_disk(ip, dip, iip->ili_item.li_lsn);
3855
3856 /* Wrap, we never let the log put out DI_MAX_FLUSH */
3857 if (ip->i_d.di_flushiter == DI_MAX_FLUSH)
3858 ip->i_d.di_flushiter = 0;
3859
3860 xfs_iflush_fork(ip, dip, iip, XFS_DATA_FORK);
3861 if (XFS_IFORK_Q(ip))
3862 xfs_iflush_fork(ip, dip, iip, XFS_ATTR_FORK);
3863 xfs_inobp_check(mp, bp);
3864
3865 /*
3866 * We've recorded everything logged in the inode, so we'd like to clear
3867 * the ili_fields bits so we don't log and flush things unnecessarily.
3868 * However, we can't stop logging all this information until the data
3869 * we've copied into the disk buffer is written to disk. If we did we
3870 * might overwrite the copy of the inode in the log with all the data
3871 * after re-logging only part of it, and in the face of a crash we
3872 * wouldn't have all the data we need to recover.
3873 *
3874 * What we do is move the bits to the ili_last_fields field. When
3875 * logging the inode, these bits are moved back to the ili_fields field.
3876 * In the xfs_iflush_done() routine we clear ili_last_fields, since we
3877 * know that the information those bits represent is permanently on
3878 * disk. As long as the flush completes before the inode is logged
3879 * again, then both ili_fields and ili_last_fields will be cleared.
3880 *
3881 * We can play with the ili_fields bits here, because the inode lock
3882 * must be held exclusively in order to set bits there and the flush
3883 * lock protects the ili_last_fields bits. Set ili_logged so the flush
3884 * done routine can tell whether or not to look in the AIL. Also, store
3885 * the current LSN of the inode so that we can tell whether the item has
3886 * moved in the AIL from xfs_iflush_done(). In order to read the lsn we
3887 * need the AIL lock, because it is a 64 bit value that cannot be read
3888 * atomically.
3889 */
3890 iip->ili_last_fields = iip->ili_fields;
3891 iip->ili_fields = 0;
3892 iip->ili_fsync_fields = 0;
3893 iip->ili_logged = 1;
3894
3895 xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
3896 &iip->ili_item.li_lsn);
3897
3898 /*
3899 * Attach the function xfs_iflush_done to the inode's
3900 * buffer. This will remove the inode from the AIL
3901 * and unlock the inode's flush lock when the inode is
3902 * completely written to disk.
3903 */
3904 xfs_buf_attach_iodone(bp, xfs_iflush_done, &iip->ili_item);
3905
3906 /* generate the checksum. */
3907 xfs_dinode_calc_crc(mp, dip);
3908
3909 ASSERT(!list_empty(&bp->b_li_list));
3910 ASSERT(bp->b_iodone != NULL);
3911 return 0;
3912
3913corrupt_out:
3914 return -EFSCORRUPTED;
3915}
3916
3917/* Release an inode. */
3918void
3919xfs_irele(
3920 struct xfs_inode *ip)
3921{
3922 trace_xfs_irele(ip, _RET_IP_);
3923 iput(VFS_I(ip));
3924}
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_ialloc.h"
24#include "xfs_bmap.h"
25#include "xfs_bmap_util.h"
26#include "xfs_errortag.h"
27#include "xfs_error.h"
28#include "xfs_quota.h"
29#include "xfs_filestream.h"
30#include "xfs_trace.h"
31#include "xfs_icache.h"
32#include "xfs_symlink.h"
33#include "xfs_trans_priv.h"
34#include "xfs_log.h"
35#include "xfs_bmap_btree.h"
36#include "xfs_reflink.h"
37#include "xfs_ag.h"
38
39kmem_zone_t *xfs_inode_zone;
40
41/*
42 * Used in xfs_itruncate_extents(). This is the maximum number of extents
43 * freed from a file in a single transaction.
44 */
45#define XFS_ITRUNC_MAX_EXTENTS 2
46
47STATIC int xfs_iunlink(struct xfs_trans *, struct xfs_inode *);
48STATIC int xfs_iunlink_remove(struct xfs_trans *tp, struct xfs_perag *pag,
49 struct xfs_inode *);
50
51/*
52 * helper function to extract extent size hint from inode
53 */
54xfs_extlen_t
55xfs_get_extsz_hint(
56 struct xfs_inode *ip)
57{
58 /*
59 * No point in aligning allocations if we need to COW to actually
60 * write to them.
61 */
62 if (xfs_is_always_cow_inode(ip))
63 return 0;
64 if ((ip->i_diflags & XFS_DIFLAG_EXTSIZE) && ip->i_extsize)
65 return ip->i_extsize;
66 if (XFS_IS_REALTIME_INODE(ip))
67 return ip->i_mount->m_sb.sb_rextsize;
68 return 0;
69}
70
71/*
72 * Helper function to extract CoW extent size hint from inode.
73 * Between the extent size hint and the CoW extent size hint, we
74 * return the greater of the two. If the value is zero (automatic),
75 * use the default size.
76 */
77xfs_extlen_t
78xfs_get_cowextsz_hint(
79 struct xfs_inode *ip)
80{
81 xfs_extlen_t a, b;
82
83 a = 0;
84 if (ip->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE)
85 a = ip->i_cowextsize;
86 b = xfs_get_extsz_hint(ip);
87
88 a = max(a, b);
89 if (a == 0)
90 return XFS_DEFAULT_COWEXTSZ_HINT;
91 return a;
92}
93
94/*
95 * These two are wrapper routines around the xfs_ilock() routine used to
96 * centralize some grungy code. They are used in places that wish to lock the
97 * inode solely for reading the extents. The reason these places can't just
98 * call xfs_ilock(ip, XFS_ILOCK_SHARED) is that the inode lock also guards to
99 * bringing in of the extents from disk for a file in b-tree format. If the
100 * inode is in b-tree format, then we need to lock the inode exclusively until
101 * the extents are read in. Locking it exclusively all the time would limit
102 * our parallelism unnecessarily, though. What we do instead is check to see
103 * if the extents have been read in yet, and only lock the inode exclusively
104 * if they have not.
105 *
106 * The functions return a value which should be given to the corresponding
107 * xfs_iunlock() call.
108 */
109uint
110xfs_ilock_data_map_shared(
111 struct xfs_inode *ip)
112{
113 uint lock_mode = XFS_ILOCK_SHARED;
114
115 if (xfs_need_iread_extents(&ip->i_df))
116 lock_mode = XFS_ILOCK_EXCL;
117 xfs_ilock(ip, lock_mode);
118 return lock_mode;
119}
120
121uint
122xfs_ilock_attr_map_shared(
123 struct xfs_inode *ip)
124{
125 uint lock_mode = XFS_ILOCK_SHARED;
126
127 if (ip->i_afp && xfs_need_iread_extents(ip->i_afp))
128 lock_mode = XFS_ILOCK_EXCL;
129 xfs_ilock(ip, lock_mode);
130 return lock_mode;
131}
132
133/*
134 * In addition to i_rwsem in the VFS inode, the xfs inode contains 2
135 * multi-reader locks: i_mmap_lock and the i_lock. This routine allows
136 * various combinations of the locks to be obtained.
137 *
138 * The 3 locks should always be ordered so that the IO lock is obtained first,
139 * the mmap lock second and the ilock last in order to prevent deadlock.
140 *
141 * Basic locking order:
142 *
143 * i_rwsem -> i_mmap_lock -> page_lock -> i_ilock
144 *
145 * mmap_lock locking order:
146 *
147 * i_rwsem -> page lock -> mmap_lock
148 * mmap_lock -> i_mmap_lock -> page_lock
149 *
150 * The difference in mmap_lock locking order mean that we cannot hold the
151 * i_mmap_lock over syscall based read(2)/write(2) based IO. These IO paths can
152 * fault in pages during copy in/out (for buffered IO) or require the mmap_lock
153 * in get_user_pages() to map the user pages into the kernel address space for
154 * direct IO. Similarly the i_rwsem cannot be taken inside a page fault because
155 * page faults already hold the mmap_lock.
156 *
157 * Hence to serialise fully against both syscall and mmap based IO, we need to
158 * take both the i_rwsem and the i_mmap_lock. These locks should *only* be both
159 * taken in places where we need to invalidate the page cache in a race
160 * free manner (e.g. truncate, hole punch and other extent manipulation
161 * functions).
162 */
163void
164xfs_ilock(
165 xfs_inode_t *ip,
166 uint lock_flags)
167{
168 trace_xfs_ilock(ip, lock_flags, _RET_IP_);
169
170 /*
171 * You can't set both SHARED and EXCL for the same lock,
172 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
173 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
174 */
175 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
176 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
177 ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
178 (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
179 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
180 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
181 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
182
183 if (lock_flags & XFS_IOLOCK_EXCL) {
184 down_write_nested(&VFS_I(ip)->i_rwsem,
185 XFS_IOLOCK_DEP(lock_flags));
186 } else if (lock_flags & XFS_IOLOCK_SHARED) {
187 down_read_nested(&VFS_I(ip)->i_rwsem,
188 XFS_IOLOCK_DEP(lock_flags));
189 }
190
191 if (lock_flags & XFS_MMAPLOCK_EXCL)
192 mrupdate_nested(&ip->i_mmaplock, XFS_MMAPLOCK_DEP(lock_flags));
193 else if (lock_flags & XFS_MMAPLOCK_SHARED)
194 mraccess_nested(&ip->i_mmaplock, XFS_MMAPLOCK_DEP(lock_flags));
195
196 if (lock_flags & XFS_ILOCK_EXCL)
197 mrupdate_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
198 else if (lock_flags & XFS_ILOCK_SHARED)
199 mraccess_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
200}
201
202/*
203 * This is just like xfs_ilock(), except that the caller
204 * is guaranteed not to sleep. It returns 1 if it gets
205 * the requested locks and 0 otherwise. If the IO lock is
206 * obtained but the inode lock cannot be, then the IO lock
207 * is dropped before returning.
208 *
209 * ip -- the inode being locked
210 * lock_flags -- this parameter indicates the inode's locks to be
211 * to be locked. See the comment for xfs_ilock() for a list
212 * of valid values.
213 */
214int
215xfs_ilock_nowait(
216 xfs_inode_t *ip,
217 uint lock_flags)
218{
219 trace_xfs_ilock_nowait(ip, lock_flags, _RET_IP_);
220
221 /*
222 * You can't set both SHARED and EXCL for the same lock,
223 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
224 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
225 */
226 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
227 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
228 ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
229 (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
230 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
231 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
232 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
233
234 if (lock_flags & XFS_IOLOCK_EXCL) {
235 if (!down_write_trylock(&VFS_I(ip)->i_rwsem))
236 goto out;
237 } else if (lock_flags & XFS_IOLOCK_SHARED) {
238 if (!down_read_trylock(&VFS_I(ip)->i_rwsem))
239 goto out;
240 }
241
242 if (lock_flags & XFS_MMAPLOCK_EXCL) {
243 if (!mrtryupdate(&ip->i_mmaplock))
244 goto out_undo_iolock;
245 } else if (lock_flags & XFS_MMAPLOCK_SHARED) {
246 if (!mrtryaccess(&ip->i_mmaplock))
247 goto out_undo_iolock;
248 }
249
250 if (lock_flags & XFS_ILOCK_EXCL) {
251 if (!mrtryupdate(&ip->i_lock))
252 goto out_undo_mmaplock;
253 } else if (lock_flags & XFS_ILOCK_SHARED) {
254 if (!mrtryaccess(&ip->i_lock))
255 goto out_undo_mmaplock;
256 }
257 return 1;
258
259out_undo_mmaplock:
260 if (lock_flags & XFS_MMAPLOCK_EXCL)
261 mrunlock_excl(&ip->i_mmaplock);
262 else if (lock_flags & XFS_MMAPLOCK_SHARED)
263 mrunlock_shared(&ip->i_mmaplock);
264out_undo_iolock:
265 if (lock_flags & XFS_IOLOCK_EXCL)
266 up_write(&VFS_I(ip)->i_rwsem);
267 else if (lock_flags & XFS_IOLOCK_SHARED)
268 up_read(&VFS_I(ip)->i_rwsem);
269out:
270 return 0;
271}
272
273/*
274 * xfs_iunlock() is used to drop the inode locks acquired with
275 * xfs_ilock() and xfs_ilock_nowait(). The caller must pass
276 * in the flags given to xfs_ilock() or xfs_ilock_nowait() so
277 * that we know which locks to drop.
278 *
279 * ip -- the inode being unlocked
280 * lock_flags -- this parameter indicates the inode's locks to be
281 * to be unlocked. See the comment for xfs_ilock() for a list
282 * of valid values for this parameter.
283 *
284 */
285void
286xfs_iunlock(
287 xfs_inode_t *ip,
288 uint lock_flags)
289{
290 /*
291 * You can't set both SHARED and EXCL for the same lock,
292 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
293 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
294 */
295 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
296 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
297 ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
298 (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
299 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
300 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
301 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
302 ASSERT(lock_flags != 0);
303
304 if (lock_flags & XFS_IOLOCK_EXCL)
305 up_write(&VFS_I(ip)->i_rwsem);
306 else if (lock_flags & XFS_IOLOCK_SHARED)
307 up_read(&VFS_I(ip)->i_rwsem);
308
309 if (lock_flags & XFS_MMAPLOCK_EXCL)
310 mrunlock_excl(&ip->i_mmaplock);
311 else if (lock_flags & XFS_MMAPLOCK_SHARED)
312 mrunlock_shared(&ip->i_mmaplock);
313
314 if (lock_flags & XFS_ILOCK_EXCL)
315 mrunlock_excl(&ip->i_lock);
316 else if (lock_flags & XFS_ILOCK_SHARED)
317 mrunlock_shared(&ip->i_lock);
318
319 trace_xfs_iunlock(ip, lock_flags, _RET_IP_);
320}
321
322/*
323 * give up write locks. the i/o lock cannot be held nested
324 * if it is being demoted.
325 */
326void
327xfs_ilock_demote(
328 xfs_inode_t *ip,
329 uint lock_flags)
330{
331 ASSERT(lock_flags & (XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL));
332 ASSERT((lock_flags &
333 ~(XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL)) == 0);
334
335 if (lock_flags & XFS_ILOCK_EXCL)
336 mrdemote(&ip->i_lock);
337 if (lock_flags & XFS_MMAPLOCK_EXCL)
338 mrdemote(&ip->i_mmaplock);
339 if (lock_flags & XFS_IOLOCK_EXCL)
340 downgrade_write(&VFS_I(ip)->i_rwsem);
341
342 trace_xfs_ilock_demote(ip, lock_flags, _RET_IP_);
343}
344
345#if defined(DEBUG) || defined(XFS_WARN)
346int
347xfs_isilocked(
348 xfs_inode_t *ip,
349 uint lock_flags)
350{
351 if (lock_flags & (XFS_ILOCK_EXCL|XFS_ILOCK_SHARED)) {
352 if (!(lock_flags & XFS_ILOCK_SHARED))
353 return !!ip->i_lock.mr_writer;
354 return rwsem_is_locked(&ip->i_lock.mr_lock);
355 }
356
357 if (lock_flags & (XFS_MMAPLOCK_EXCL|XFS_MMAPLOCK_SHARED)) {
358 if (!(lock_flags & XFS_MMAPLOCK_SHARED))
359 return !!ip->i_mmaplock.mr_writer;
360 return rwsem_is_locked(&ip->i_mmaplock.mr_lock);
361 }
362
363 if (lock_flags & (XFS_IOLOCK_EXCL|XFS_IOLOCK_SHARED)) {
364 if (!(lock_flags & XFS_IOLOCK_SHARED))
365 return !debug_locks ||
366 lockdep_is_held_type(&VFS_I(ip)->i_rwsem, 0);
367 return rwsem_is_locked(&VFS_I(ip)->i_rwsem);
368 }
369
370 ASSERT(0);
371 return 0;
372}
373#endif
374
375/*
376 * xfs_lockdep_subclass_ok() is only used in an ASSERT, so is only called when
377 * DEBUG or XFS_WARN is set. And MAX_LOCKDEP_SUBCLASSES is then only defined
378 * when CONFIG_LOCKDEP is set. Hence the complex define below to avoid build
379 * errors and warnings.
380 */
381#if (defined(DEBUG) || defined(XFS_WARN)) && defined(CONFIG_LOCKDEP)
382static bool
383xfs_lockdep_subclass_ok(
384 int subclass)
385{
386 return subclass < MAX_LOCKDEP_SUBCLASSES;
387}
388#else
389#define xfs_lockdep_subclass_ok(subclass) (true)
390#endif
391
392/*
393 * Bump the subclass so xfs_lock_inodes() acquires each lock with a different
394 * value. This can be called for any type of inode lock combination, including
395 * parent locking. Care must be taken to ensure we don't overrun the subclass
396 * storage fields in the class mask we build.
397 */
398static inline int
399xfs_lock_inumorder(int lock_mode, int subclass)
400{
401 int class = 0;
402
403 ASSERT(!(lock_mode & (XFS_ILOCK_PARENT | XFS_ILOCK_RTBITMAP |
404 XFS_ILOCK_RTSUM)));
405 ASSERT(xfs_lockdep_subclass_ok(subclass));
406
407 if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)) {
408 ASSERT(subclass <= XFS_IOLOCK_MAX_SUBCLASS);
409 class += subclass << XFS_IOLOCK_SHIFT;
410 }
411
412 if (lock_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) {
413 ASSERT(subclass <= XFS_MMAPLOCK_MAX_SUBCLASS);
414 class += subclass << XFS_MMAPLOCK_SHIFT;
415 }
416
417 if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) {
418 ASSERT(subclass <= XFS_ILOCK_MAX_SUBCLASS);
419 class += subclass << XFS_ILOCK_SHIFT;
420 }
421
422 return (lock_mode & ~XFS_LOCK_SUBCLASS_MASK) | class;
423}
424
425/*
426 * The following routine will lock n inodes in exclusive mode. We assume the
427 * caller calls us with the inodes in i_ino order.
428 *
429 * We need to detect deadlock where an inode that we lock is in the AIL and we
430 * start waiting for another inode that is locked by a thread in a long running
431 * transaction (such as truncate). This can result in deadlock since the long
432 * running trans might need to wait for the inode we just locked in order to
433 * push the tail and free space in the log.
434 *
435 * xfs_lock_inodes() can only be used to lock one type of lock at a time -
436 * the iolock, the mmaplock or the ilock, but not more than one at a time. If we
437 * lock more than one at a time, lockdep will report false positives saying we
438 * have violated locking orders.
439 */
440static void
441xfs_lock_inodes(
442 struct xfs_inode **ips,
443 int inodes,
444 uint lock_mode)
445{
446 int attempts = 0, i, j, try_lock;
447 struct xfs_log_item *lp;
448
449 /*
450 * Currently supports between 2 and 5 inodes with exclusive locking. We
451 * support an arbitrary depth of locking here, but absolute limits on
452 * inodes depend on the type of locking and the limits placed by
453 * lockdep annotations in xfs_lock_inumorder. These are all checked by
454 * the asserts.
455 */
456 ASSERT(ips && inodes >= 2 && inodes <= 5);
457 ASSERT(lock_mode & (XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL |
458 XFS_ILOCK_EXCL));
459 ASSERT(!(lock_mode & (XFS_IOLOCK_SHARED | XFS_MMAPLOCK_SHARED |
460 XFS_ILOCK_SHARED)));
461 ASSERT(!(lock_mode & XFS_MMAPLOCK_EXCL) ||
462 inodes <= XFS_MMAPLOCK_MAX_SUBCLASS + 1);
463 ASSERT(!(lock_mode & XFS_ILOCK_EXCL) ||
464 inodes <= XFS_ILOCK_MAX_SUBCLASS + 1);
465
466 if (lock_mode & XFS_IOLOCK_EXCL) {
467 ASSERT(!(lock_mode & (XFS_MMAPLOCK_EXCL | XFS_ILOCK_EXCL)));
468 } else if (lock_mode & XFS_MMAPLOCK_EXCL)
469 ASSERT(!(lock_mode & XFS_ILOCK_EXCL));
470
471 try_lock = 0;
472 i = 0;
473again:
474 for (; i < inodes; i++) {
475 ASSERT(ips[i]);
476
477 if (i && (ips[i] == ips[i - 1])) /* Already locked */
478 continue;
479
480 /*
481 * If try_lock is not set yet, make sure all locked inodes are
482 * not in the AIL. If any are, set try_lock to be used later.
483 */
484 if (!try_lock) {
485 for (j = (i - 1); j >= 0 && !try_lock; j--) {
486 lp = &ips[j]->i_itemp->ili_item;
487 if (lp && test_bit(XFS_LI_IN_AIL, &lp->li_flags))
488 try_lock++;
489 }
490 }
491
492 /*
493 * If any of the previous locks we have locked is in the AIL,
494 * we must TRY to get the second and subsequent locks. If
495 * we can't get any, we must release all we have
496 * and try again.
497 */
498 if (!try_lock) {
499 xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
500 continue;
501 }
502
503 /* try_lock means we have an inode locked that is in the AIL. */
504 ASSERT(i != 0);
505 if (xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i)))
506 continue;
507
508 /*
509 * Unlock all previous guys and try again. xfs_iunlock will try
510 * to push the tail if the inode is in the AIL.
511 */
512 attempts++;
513 for (j = i - 1; j >= 0; j--) {
514 /*
515 * Check to see if we've already unlocked this one. Not
516 * the first one going back, and the inode ptr is the
517 * same.
518 */
519 if (j != (i - 1) && ips[j] == ips[j + 1])
520 continue;
521
522 xfs_iunlock(ips[j], lock_mode);
523 }
524
525 if ((attempts % 5) == 0) {
526 delay(1); /* Don't just spin the CPU */
527 }
528 i = 0;
529 try_lock = 0;
530 goto again;
531 }
532}
533
534/*
535 * xfs_lock_two_inodes() can only be used to lock one type of lock at a time -
536 * the mmaplock or the ilock, but not more than one type at a time. If we lock
537 * more than one at a time, lockdep will report false positives saying we have
538 * violated locking orders. The iolock must be double-locked separately since
539 * we use i_rwsem for that. We now support taking one lock EXCL and the other
540 * SHARED.
541 */
542void
543xfs_lock_two_inodes(
544 struct xfs_inode *ip0,
545 uint ip0_mode,
546 struct xfs_inode *ip1,
547 uint ip1_mode)
548{
549 struct xfs_inode *temp;
550 uint mode_temp;
551 int attempts = 0;
552 struct xfs_log_item *lp;
553
554 ASSERT(hweight32(ip0_mode) == 1);
555 ASSERT(hweight32(ip1_mode) == 1);
556 ASSERT(!(ip0_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
557 ASSERT(!(ip1_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
558 ASSERT(!(ip0_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) ||
559 !(ip0_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
560 ASSERT(!(ip1_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) ||
561 !(ip1_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
562 ASSERT(!(ip1_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) ||
563 !(ip0_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
564 ASSERT(!(ip0_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) ||
565 !(ip1_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
566
567 ASSERT(ip0->i_ino != ip1->i_ino);
568
569 if (ip0->i_ino > ip1->i_ino) {
570 temp = ip0;
571 ip0 = ip1;
572 ip1 = temp;
573 mode_temp = ip0_mode;
574 ip0_mode = ip1_mode;
575 ip1_mode = mode_temp;
576 }
577
578 again:
579 xfs_ilock(ip0, xfs_lock_inumorder(ip0_mode, 0));
580
581 /*
582 * If the first lock we have locked is in the AIL, we must TRY to get
583 * the second lock. If we can't get it, we must release the first one
584 * and try again.
585 */
586 lp = &ip0->i_itemp->ili_item;
587 if (lp && test_bit(XFS_LI_IN_AIL, &lp->li_flags)) {
588 if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(ip1_mode, 1))) {
589 xfs_iunlock(ip0, ip0_mode);
590 if ((++attempts % 5) == 0)
591 delay(1); /* Don't just spin the CPU */
592 goto again;
593 }
594 } else {
595 xfs_ilock(ip1, xfs_lock_inumorder(ip1_mode, 1));
596 }
597}
598
599uint
600xfs_ip2xflags(
601 struct xfs_inode *ip)
602{
603 uint flags = 0;
604
605 if (ip->i_diflags & XFS_DIFLAG_ANY) {
606 if (ip->i_diflags & XFS_DIFLAG_REALTIME)
607 flags |= FS_XFLAG_REALTIME;
608 if (ip->i_diflags & XFS_DIFLAG_PREALLOC)
609 flags |= FS_XFLAG_PREALLOC;
610 if (ip->i_diflags & XFS_DIFLAG_IMMUTABLE)
611 flags |= FS_XFLAG_IMMUTABLE;
612 if (ip->i_diflags & XFS_DIFLAG_APPEND)
613 flags |= FS_XFLAG_APPEND;
614 if (ip->i_diflags & XFS_DIFLAG_SYNC)
615 flags |= FS_XFLAG_SYNC;
616 if (ip->i_diflags & XFS_DIFLAG_NOATIME)
617 flags |= FS_XFLAG_NOATIME;
618 if (ip->i_diflags & XFS_DIFLAG_NODUMP)
619 flags |= FS_XFLAG_NODUMP;
620 if (ip->i_diflags & XFS_DIFLAG_RTINHERIT)
621 flags |= FS_XFLAG_RTINHERIT;
622 if (ip->i_diflags & XFS_DIFLAG_PROJINHERIT)
623 flags |= FS_XFLAG_PROJINHERIT;
624 if (ip->i_diflags & XFS_DIFLAG_NOSYMLINKS)
625 flags |= FS_XFLAG_NOSYMLINKS;
626 if (ip->i_diflags & XFS_DIFLAG_EXTSIZE)
627 flags |= FS_XFLAG_EXTSIZE;
628 if (ip->i_diflags & XFS_DIFLAG_EXTSZINHERIT)
629 flags |= FS_XFLAG_EXTSZINHERIT;
630 if (ip->i_diflags & XFS_DIFLAG_NODEFRAG)
631 flags |= FS_XFLAG_NODEFRAG;
632 if (ip->i_diflags & XFS_DIFLAG_FILESTREAM)
633 flags |= FS_XFLAG_FILESTREAM;
634 }
635
636 if (ip->i_diflags2 & XFS_DIFLAG2_ANY) {
637 if (ip->i_diflags2 & XFS_DIFLAG2_DAX)
638 flags |= FS_XFLAG_DAX;
639 if (ip->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE)
640 flags |= FS_XFLAG_COWEXTSIZE;
641 }
642
643 if (XFS_IFORK_Q(ip))
644 flags |= FS_XFLAG_HASATTR;
645 return flags;
646}
647
648/*
649 * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
650 * is allowed, otherwise it has to be an exact match. If a CI match is found,
651 * ci_name->name will point to a the actual name (caller must free) or
652 * will be set to NULL if an exact match is found.
653 */
654int
655xfs_lookup(
656 xfs_inode_t *dp,
657 struct xfs_name *name,
658 xfs_inode_t **ipp,
659 struct xfs_name *ci_name)
660{
661 xfs_ino_t inum;
662 int error;
663
664 trace_xfs_lookup(dp, name);
665
666 if (XFS_FORCED_SHUTDOWN(dp->i_mount))
667 return -EIO;
668
669 error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
670 if (error)
671 goto out_unlock;
672
673 error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp);
674 if (error)
675 goto out_free_name;
676
677 return 0;
678
679out_free_name:
680 if (ci_name)
681 kmem_free(ci_name->name);
682out_unlock:
683 *ipp = NULL;
684 return error;
685}
686
687/* Propagate di_flags from a parent inode to a child inode. */
688static void
689xfs_inode_inherit_flags(
690 struct xfs_inode *ip,
691 const struct xfs_inode *pip)
692{
693 unsigned int di_flags = 0;
694 xfs_failaddr_t failaddr;
695 umode_t mode = VFS_I(ip)->i_mode;
696
697 if (S_ISDIR(mode)) {
698 if (pip->i_diflags & XFS_DIFLAG_RTINHERIT)
699 di_flags |= XFS_DIFLAG_RTINHERIT;
700 if (pip->i_diflags & XFS_DIFLAG_EXTSZINHERIT) {
701 di_flags |= XFS_DIFLAG_EXTSZINHERIT;
702 ip->i_extsize = pip->i_extsize;
703 }
704 if (pip->i_diflags & XFS_DIFLAG_PROJINHERIT)
705 di_flags |= XFS_DIFLAG_PROJINHERIT;
706 } else if (S_ISREG(mode)) {
707 if ((pip->i_diflags & XFS_DIFLAG_RTINHERIT) &&
708 xfs_sb_version_hasrealtime(&ip->i_mount->m_sb))
709 di_flags |= XFS_DIFLAG_REALTIME;
710 if (pip->i_diflags & XFS_DIFLAG_EXTSZINHERIT) {
711 di_flags |= XFS_DIFLAG_EXTSIZE;
712 ip->i_extsize = pip->i_extsize;
713 }
714 }
715 if ((pip->i_diflags & XFS_DIFLAG_NOATIME) &&
716 xfs_inherit_noatime)
717 di_flags |= XFS_DIFLAG_NOATIME;
718 if ((pip->i_diflags & XFS_DIFLAG_NODUMP) &&
719 xfs_inherit_nodump)
720 di_flags |= XFS_DIFLAG_NODUMP;
721 if ((pip->i_diflags & XFS_DIFLAG_SYNC) &&
722 xfs_inherit_sync)
723 di_flags |= XFS_DIFLAG_SYNC;
724 if ((pip->i_diflags & XFS_DIFLAG_NOSYMLINKS) &&
725 xfs_inherit_nosymlinks)
726 di_flags |= XFS_DIFLAG_NOSYMLINKS;
727 if ((pip->i_diflags & XFS_DIFLAG_NODEFRAG) &&
728 xfs_inherit_nodefrag)
729 di_flags |= XFS_DIFLAG_NODEFRAG;
730 if (pip->i_diflags & XFS_DIFLAG_FILESTREAM)
731 di_flags |= XFS_DIFLAG_FILESTREAM;
732
733 ip->i_diflags |= di_flags;
734
735 /*
736 * Inode verifiers on older kernels only check that the extent size
737 * hint is an integer multiple of the rt extent size on realtime files.
738 * They did not check the hint alignment on a directory with both
739 * rtinherit and extszinherit flags set. If the misaligned hint is
740 * propagated from a directory into a new realtime file, new file
741 * allocations will fail due to math errors in the rt allocator and/or
742 * trip the verifiers. Validate the hint settings in the new file so
743 * that we don't let broken hints propagate.
744 */
745 failaddr = xfs_inode_validate_extsize(ip->i_mount, ip->i_extsize,
746 VFS_I(ip)->i_mode, ip->i_diflags);
747 if (failaddr) {
748 ip->i_diflags &= ~(XFS_DIFLAG_EXTSIZE |
749 XFS_DIFLAG_EXTSZINHERIT);
750 ip->i_extsize = 0;
751 }
752}
753
754/* Propagate di_flags2 from a parent inode to a child inode. */
755static void
756xfs_inode_inherit_flags2(
757 struct xfs_inode *ip,
758 const struct xfs_inode *pip)
759{
760 xfs_failaddr_t failaddr;
761
762 if (pip->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE) {
763 ip->i_diflags2 |= XFS_DIFLAG2_COWEXTSIZE;
764 ip->i_cowextsize = pip->i_cowextsize;
765 }
766 if (pip->i_diflags2 & XFS_DIFLAG2_DAX)
767 ip->i_diflags2 |= XFS_DIFLAG2_DAX;
768
769 /* Don't let invalid cowextsize hints propagate. */
770 failaddr = xfs_inode_validate_cowextsize(ip->i_mount, ip->i_cowextsize,
771 VFS_I(ip)->i_mode, ip->i_diflags, ip->i_diflags2);
772 if (failaddr) {
773 ip->i_diflags2 &= ~XFS_DIFLAG2_COWEXTSIZE;
774 ip->i_cowextsize = 0;
775 }
776}
777
778/*
779 * Initialise a newly allocated inode and return the in-core inode to the
780 * caller locked exclusively.
781 */
782int
783xfs_init_new_inode(
784 struct user_namespace *mnt_userns,
785 struct xfs_trans *tp,
786 struct xfs_inode *pip,
787 xfs_ino_t ino,
788 umode_t mode,
789 xfs_nlink_t nlink,
790 dev_t rdev,
791 prid_t prid,
792 bool init_xattrs,
793 struct xfs_inode **ipp)
794{
795 struct inode *dir = pip ? VFS_I(pip) : NULL;
796 struct xfs_mount *mp = tp->t_mountp;
797 struct xfs_inode *ip;
798 unsigned int flags;
799 int error;
800 struct timespec64 tv;
801 struct inode *inode;
802
803 /*
804 * Protect against obviously corrupt allocation btree records. Later
805 * xfs_iget checks will catch re-allocation of other active in-memory
806 * and on-disk inodes. If we don't catch reallocating the parent inode
807 * here we will deadlock in xfs_iget() so we have to do these checks
808 * first.
809 */
810 if ((pip && ino == pip->i_ino) || !xfs_verify_dir_ino(mp, ino)) {
811 xfs_alert(mp, "Allocated a known in-use inode 0x%llx!", ino);
812 return -EFSCORRUPTED;
813 }
814
815 /*
816 * Get the in-core inode with the lock held exclusively to prevent
817 * others from looking at until we're done.
818 */
819 error = xfs_iget(mp, tp, ino, XFS_IGET_CREATE, XFS_ILOCK_EXCL, &ip);
820 if (error)
821 return error;
822
823 ASSERT(ip != NULL);
824 inode = VFS_I(ip);
825 set_nlink(inode, nlink);
826 inode->i_rdev = rdev;
827 ip->i_projid = prid;
828
829 if (dir && !(dir->i_mode & S_ISGID) &&
830 (mp->m_flags & XFS_MOUNT_GRPID)) {
831 inode_fsuid_set(inode, mnt_userns);
832 inode->i_gid = dir->i_gid;
833 inode->i_mode = mode;
834 } else {
835 inode_init_owner(mnt_userns, inode, dir, mode);
836 }
837
838 /*
839 * If the group ID of the new file does not match the effective group
840 * ID or one of the supplementary group IDs, the S_ISGID bit is cleared
841 * (and only if the irix_sgid_inherit compatibility variable is set).
842 */
843 if (irix_sgid_inherit &&
844 (inode->i_mode & S_ISGID) &&
845 !in_group_p(i_gid_into_mnt(mnt_userns, inode)))
846 inode->i_mode &= ~S_ISGID;
847
848 ip->i_disk_size = 0;
849 ip->i_df.if_nextents = 0;
850 ASSERT(ip->i_nblocks == 0);
851
852 tv = current_time(inode);
853 inode->i_mtime = tv;
854 inode->i_atime = tv;
855 inode->i_ctime = tv;
856
857 ip->i_extsize = 0;
858 ip->i_diflags = 0;
859
860 if (xfs_sb_version_has_v3inode(&mp->m_sb)) {
861 inode_set_iversion(inode, 1);
862 ip->i_cowextsize = 0;
863 ip->i_crtime = tv;
864 }
865
866 flags = XFS_ILOG_CORE;
867 switch (mode & S_IFMT) {
868 case S_IFIFO:
869 case S_IFCHR:
870 case S_IFBLK:
871 case S_IFSOCK:
872 ip->i_df.if_format = XFS_DINODE_FMT_DEV;
873 flags |= XFS_ILOG_DEV;
874 break;
875 case S_IFREG:
876 case S_IFDIR:
877 if (pip && (pip->i_diflags & XFS_DIFLAG_ANY))
878 xfs_inode_inherit_flags(ip, pip);
879 if (pip && (pip->i_diflags2 & XFS_DIFLAG2_ANY))
880 xfs_inode_inherit_flags2(ip, pip);
881 fallthrough;
882 case S_IFLNK:
883 ip->i_df.if_format = XFS_DINODE_FMT_EXTENTS;
884 ip->i_df.if_bytes = 0;
885 ip->i_df.if_u1.if_root = NULL;
886 break;
887 default:
888 ASSERT(0);
889 }
890
891 /*
892 * If we need to create attributes immediately after allocating the
893 * inode, initialise an empty attribute fork right now. We use the
894 * default fork offset for attributes here as we don't know exactly what
895 * size or how many attributes we might be adding. We can do this
896 * safely here because we know the data fork is completely empty and
897 * this saves us from needing to run a separate transaction to set the
898 * fork offset in the immediate future.
899 */
900 if (init_xattrs && xfs_sb_version_hasattr(&mp->m_sb)) {
901 ip->i_forkoff = xfs_default_attroffset(ip) >> 3;
902 ip->i_afp = xfs_ifork_alloc(XFS_DINODE_FMT_EXTENTS, 0);
903 }
904
905 /*
906 * Log the new values stuffed into the inode.
907 */
908 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
909 xfs_trans_log_inode(tp, ip, flags);
910
911 /* now that we have an i_mode we can setup the inode structure */
912 xfs_setup_inode(ip);
913
914 *ipp = ip;
915 return 0;
916}
917
918/*
919 * Decrement the link count on an inode & log the change. If this causes the
920 * link count to go to zero, move the inode to AGI unlinked list so that it can
921 * be freed when the last active reference goes away via xfs_inactive().
922 */
923static int /* error */
924xfs_droplink(
925 xfs_trans_t *tp,
926 xfs_inode_t *ip)
927{
928 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
929
930 drop_nlink(VFS_I(ip));
931 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
932
933 if (VFS_I(ip)->i_nlink)
934 return 0;
935
936 return xfs_iunlink(tp, ip);
937}
938
939/*
940 * Increment the link count on an inode & log the change.
941 */
942static void
943xfs_bumplink(
944 xfs_trans_t *tp,
945 xfs_inode_t *ip)
946{
947 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
948
949 inc_nlink(VFS_I(ip));
950 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
951}
952
953int
954xfs_create(
955 struct user_namespace *mnt_userns,
956 xfs_inode_t *dp,
957 struct xfs_name *name,
958 umode_t mode,
959 dev_t rdev,
960 bool init_xattrs,
961 xfs_inode_t **ipp)
962{
963 int is_dir = S_ISDIR(mode);
964 struct xfs_mount *mp = dp->i_mount;
965 struct xfs_inode *ip = NULL;
966 struct xfs_trans *tp = NULL;
967 int error;
968 bool unlock_dp_on_error = false;
969 prid_t prid;
970 struct xfs_dquot *udqp = NULL;
971 struct xfs_dquot *gdqp = NULL;
972 struct xfs_dquot *pdqp = NULL;
973 struct xfs_trans_res *tres;
974 uint resblks;
975 xfs_ino_t ino;
976
977 trace_xfs_create(dp, name);
978
979 if (XFS_FORCED_SHUTDOWN(mp))
980 return -EIO;
981
982 prid = xfs_get_initial_prid(dp);
983
984 /*
985 * Make sure that we have allocated dquot(s) on disk.
986 */
987 error = xfs_qm_vop_dqalloc(dp, mapped_fsuid(mnt_userns),
988 mapped_fsgid(mnt_userns), prid,
989 XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
990 &udqp, &gdqp, &pdqp);
991 if (error)
992 return error;
993
994 if (is_dir) {
995 resblks = XFS_MKDIR_SPACE_RES(mp, name->len);
996 tres = &M_RES(mp)->tr_mkdir;
997 } else {
998 resblks = XFS_CREATE_SPACE_RES(mp, name->len);
999 tres = &M_RES(mp)->tr_create;
1000 }
1001
1002 /*
1003 * Initially assume that the file does not exist and
1004 * reserve the resources for that case. If that is not
1005 * the case we'll drop the one we have and get a more
1006 * appropriate transaction later.
1007 */
1008 error = xfs_trans_alloc_icreate(mp, tres, udqp, gdqp, pdqp, resblks,
1009 &tp);
1010 if (error == -ENOSPC) {
1011 /* flush outstanding delalloc blocks and retry */
1012 xfs_flush_inodes(mp);
1013 error = xfs_trans_alloc_icreate(mp, tres, udqp, gdqp, pdqp,
1014 resblks, &tp);
1015 }
1016 if (error)
1017 goto out_release_dquots;
1018
1019 xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
1020 unlock_dp_on_error = true;
1021
1022 error = xfs_iext_count_may_overflow(dp, XFS_DATA_FORK,
1023 XFS_IEXT_DIR_MANIP_CNT(mp));
1024 if (error)
1025 goto out_trans_cancel;
1026
1027 /*
1028 * A newly created regular or special file just has one directory
1029 * entry pointing to them, but a directory also the "." entry
1030 * pointing to itself.
1031 */
1032 error = xfs_dialloc(&tp, dp->i_ino, mode, &ino);
1033 if (!error)
1034 error = xfs_init_new_inode(mnt_userns, tp, dp, ino, mode,
1035 is_dir ? 2 : 1, rdev, prid, init_xattrs, &ip);
1036 if (error)
1037 goto out_trans_cancel;
1038
1039 /*
1040 * Now we join the directory inode to the transaction. We do not do it
1041 * earlier because xfs_dialloc might commit the previous transaction
1042 * (and release all the locks). An error from here on will result in
1043 * the transaction cancel unlocking dp so don't do it explicitly in the
1044 * error path.
1045 */
1046 xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1047 unlock_dp_on_error = false;
1048
1049 error = xfs_dir_createname(tp, dp, name, ip->i_ino,
1050 resblks - XFS_IALLOC_SPACE_RES(mp));
1051 if (error) {
1052 ASSERT(error != -ENOSPC);
1053 goto out_trans_cancel;
1054 }
1055 xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1056 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1057
1058 if (is_dir) {
1059 error = xfs_dir_init(tp, ip, dp);
1060 if (error)
1061 goto out_trans_cancel;
1062
1063 xfs_bumplink(tp, dp);
1064 }
1065
1066 /*
1067 * If this is a synchronous mount, make sure that the
1068 * create transaction goes to disk before returning to
1069 * the user.
1070 */
1071 if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1072 xfs_trans_set_sync(tp);
1073
1074 /*
1075 * Attach the dquot(s) to the inodes and modify them incore.
1076 * These ids of the inode couldn't have changed since the new
1077 * inode has been locked ever since it was created.
1078 */
1079 xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1080
1081 error = xfs_trans_commit(tp);
1082 if (error)
1083 goto out_release_inode;
1084
1085 xfs_qm_dqrele(udqp);
1086 xfs_qm_dqrele(gdqp);
1087 xfs_qm_dqrele(pdqp);
1088
1089 *ipp = ip;
1090 return 0;
1091
1092 out_trans_cancel:
1093 xfs_trans_cancel(tp);
1094 out_release_inode:
1095 /*
1096 * Wait until after the current transaction is aborted to finish the
1097 * setup of the inode and release the inode. This prevents recursive
1098 * transactions and deadlocks from xfs_inactive.
1099 */
1100 if (ip) {
1101 xfs_finish_inode_setup(ip);
1102 xfs_irele(ip);
1103 }
1104 out_release_dquots:
1105 xfs_qm_dqrele(udqp);
1106 xfs_qm_dqrele(gdqp);
1107 xfs_qm_dqrele(pdqp);
1108
1109 if (unlock_dp_on_error)
1110 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1111 return error;
1112}
1113
1114int
1115xfs_create_tmpfile(
1116 struct user_namespace *mnt_userns,
1117 struct xfs_inode *dp,
1118 umode_t mode,
1119 struct xfs_inode **ipp)
1120{
1121 struct xfs_mount *mp = dp->i_mount;
1122 struct xfs_inode *ip = NULL;
1123 struct xfs_trans *tp = NULL;
1124 int error;
1125 prid_t prid;
1126 struct xfs_dquot *udqp = NULL;
1127 struct xfs_dquot *gdqp = NULL;
1128 struct xfs_dquot *pdqp = NULL;
1129 struct xfs_trans_res *tres;
1130 uint resblks;
1131 xfs_ino_t ino;
1132
1133 if (XFS_FORCED_SHUTDOWN(mp))
1134 return -EIO;
1135
1136 prid = xfs_get_initial_prid(dp);
1137
1138 /*
1139 * Make sure that we have allocated dquot(s) on disk.
1140 */
1141 error = xfs_qm_vop_dqalloc(dp, mapped_fsuid(mnt_userns),
1142 mapped_fsgid(mnt_userns), prid,
1143 XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
1144 &udqp, &gdqp, &pdqp);
1145 if (error)
1146 return error;
1147
1148 resblks = XFS_IALLOC_SPACE_RES(mp);
1149 tres = &M_RES(mp)->tr_create_tmpfile;
1150
1151 error = xfs_trans_alloc_icreate(mp, tres, udqp, gdqp, pdqp, resblks,
1152 &tp);
1153 if (error)
1154 goto out_release_dquots;
1155
1156 error = xfs_dialloc(&tp, dp->i_ino, mode, &ino);
1157 if (!error)
1158 error = xfs_init_new_inode(mnt_userns, tp, dp, ino, mode,
1159 0, 0, prid, false, &ip);
1160 if (error)
1161 goto out_trans_cancel;
1162
1163 if (mp->m_flags & XFS_MOUNT_WSYNC)
1164 xfs_trans_set_sync(tp);
1165
1166 /*
1167 * Attach the dquot(s) to the inodes and modify them incore.
1168 * These ids of the inode couldn't have changed since the new
1169 * inode has been locked ever since it was created.
1170 */
1171 xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1172
1173 error = xfs_iunlink(tp, ip);
1174 if (error)
1175 goto out_trans_cancel;
1176
1177 error = xfs_trans_commit(tp);
1178 if (error)
1179 goto out_release_inode;
1180
1181 xfs_qm_dqrele(udqp);
1182 xfs_qm_dqrele(gdqp);
1183 xfs_qm_dqrele(pdqp);
1184
1185 *ipp = ip;
1186 return 0;
1187
1188 out_trans_cancel:
1189 xfs_trans_cancel(tp);
1190 out_release_inode:
1191 /*
1192 * Wait until after the current transaction is aborted to finish the
1193 * setup of the inode and release the inode. This prevents recursive
1194 * transactions and deadlocks from xfs_inactive.
1195 */
1196 if (ip) {
1197 xfs_finish_inode_setup(ip);
1198 xfs_irele(ip);
1199 }
1200 out_release_dquots:
1201 xfs_qm_dqrele(udqp);
1202 xfs_qm_dqrele(gdqp);
1203 xfs_qm_dqrele(pdqp);
1204
1205 return error;
1206}
1207
1208int
1209xfs_link(
1210 xfs_inode_t *tdp,
1211 xfs_inode_t *sip,
1212 struct xfs_name *target_name)
1213{
1214 xfs_mount_t *mp = tdp->i_mount;
1215 xfs_trans_t *tp;
1216 int error;
1217 int resblks;
1218
1219 trace_xfs_link(tdp, target_name);
1220
1221 ASSERT(!S_ISDIR(VFS_I(sip)->i_mode));
1222
1223 if (XFS_FORCED_SHUTDOWN(mp))
1224 return -EIO;
1225
1226 error = xfs_qm_dqattach(sip);
1227 if (error)
1228 goto std_return;
1229
1230 error = xfs_qm_dqattach(tdp);
1231 if (error)
1232 goto std_return;
1233
1234 resblks = XFS_LINK_SPACE_RES(mp, target_name->len);
1235 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_link, resblks, 0, 0, &tp);
1236 if (error == -ENOSPC) {
1237 resblks = 0;
1238 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_link, 0, 0, 0, &tp);
1239 }
1240 if (error)
1241 goto std_return;
1242
1243 xfs_lock_two_inodes(sip, XFS_ILOCK_EXCL, tdp, XFS_ILOCK_EXCL);
1244
1245 xfs_trans_ijoin(tp, sip, XFS_ILOCK_EXCL);
1246 xfs_trans_ijoin(tp, tdp, XFS_ILOCK_EXCL);
1247
1248 error = xfs_iext_count_may_overflow(tdp, XFS_DATA_FORK,
1249 XFS_IEXT_DIR_MANIP_CNT(mp));
1250 if (error)
1251 goto error_return;
1252
1253 /*
1254 * If we are using project inheritance, we only allow hard link
1255 * creation in our tree when the project IDs are the same; else
1256 * the tree quota mechanism could be circumvented.
1257 */
1258 if (unlikely((tdp->i_diflags & XFS_DIFLAG_PROJINHERIT) &&
1259 tdp->i_projid != sip->i_projid)) {
1260 error = -EXDEV;
1261 goto error_return;
1262 }
1263
1264 if (!resblks) {
1265 error = xfs_dir_canenter(tp, tdp, target_name);
1266 if (error)
1267 goto error_return;
1268 }
1269
1270 /*
1271 * Handle initial link state of O_TMPFILE inode
1272 */
1273 if (VFS_I(sip)->i_nlink == 0) {
1274 struct xfs_perag *pag;
1275
1276 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, sip->i_ino));
1277 error = xfs_iunlink_remove(tp, pag, sip);
1278 xfs_perag_put(pag);
1279 if (error)
1280 goto error_return;
1281 }
1282
1283 error = xfs_dir_createname(tp, tdp, target_name, sip->i_ino,
1284 resblks);
1285 if (error)
1286 goto error_return;
1287 xfs_trans_ichgtime(tp, tdp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1288 xfs_trans_log_inode(tp, tdp, XFS_ILOG_CORE);
1289
1290 xfs_bumplink(tp, sip);
1291
1292 /*
1293 * If this is a synchronous mount, make sure that the
1294 * link transaction goes to disk before returning to
1295 * the user.
1296 */
1297 if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1298 xfs_trans_set_sync(tp);
1299
1300 return xfs_trans_commit(tp);
1301
1302 error_return:
1303 xfs_trans_cancel(tp);
1304 std_return:
1305 return error;
1306}
1307
1308/* Clear the reflink flag and the cowblocks tag if possible. */
1309static void
1310xfs_itruncate_clear_reflink_flags(
1311 struct xfs_inode *ip)
1312{
1313 struct xfs_ifork *dfork;
1314 struct xfs_ifork *cfork;
1315
1316 if (!xfs_is_reflink_inode(ip))
1317 return;
1318 dfork = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1319 cfork = XFS_IFORK_PTR(ip, XFS_COW_FORK);
1320 if (dfork->if_bytes == 0 && cfork->if_bytes == 0)
1321 ip->i_diflags2 &= ~XFS_DIFLAG2_REFLINK;
1322 if (cfork->if_bytes == 0)
1323 xfs_inode_clear_cowblocks_tag(ip);
1324}
1325
1326/*
1327 * Free up the underlying blocks past new_size. The new size must be smaller
1328 * than the current size. This routine can be used both for the attribute and
1329 * data fork, and does not modify the inode size, which is left to the caller.
1330 *
1331 * The transaction passed to this routine must have made a permanent log
1332 * reservation of at least XFS_ITRUNCATE_LOG_RES. This routine may commit the
1333 * given transaction and start new ones, so make sure everything involved in
1334 * the transaction is tidy before calling here. Some transaction will be
1335 * returned to the caller to be committed. The incoming transaction must
1336 * already include the inode, and both inode locks must be held exclusively.
1337 * The inode must also be "held" within the transaction. On return the inode
1338 * will be "held" within the returned transaction. This routine does NOT
1339 * require any disk space to be reserved for it within the transaction.
1340 *
1341 * If we get an error, we must return with the inode locked and linked into the
1342 * current transaction. This keeps things simple for the higher level code,
1343 * because it always knows that the inode is locked and held in the transaction
1344 * that returns to it whether errors occur or not. We don't mark the inode
1345 * dirty on error so that transactions can be easily aborted if possible.
1346 */
1347int
1348xfs_itruncate_extents_flags(
1349 struct xfs_trans **tpp,
1350 struct xfs_inode *ip,
1351 int whichfork,
1352 xfs_fsize_t new_size,
1353 int flags)
1354{
1355 struct xfs_mount *mp = ip->i_mount;
1356 struct xfs_trans *tp = *tpp;
1357 xfs_fileoff_t first_unmap_block;
1358 xfs_filblks_t unmap_len;
1359 int error = 0;
1360
1361 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1362 ASSERT(!atomic_read(&VFS_I(ip)->i_count) ||
1363 xfs_isilocked(ip, XFS_IOLOCK_EXCL));
1364 ASSERT(new_size <= XFS_ISIZE(ip));
1365 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
1366 ASSERT(ip->i_itemp != NULL);
1367 ASSERT(ip->i_itemp->ili_lock_flags == 0);
1368 ASSERT(!XFS_NOT_DQATTACHED(mp, ip));
1369
1370 trace_xfs_itruncate_extents_start(ip, new_size);
1371
1372 flags |= xfs_bmapi_aflag(whichfork);
1373
1374 /*
1375 * Since it is possible for space to become allocated beyond
1376 * the end of the file (in a crash where the space is allocated
1377 * but the inode size is not yet updated), simply remove any
1378 * blocks which show up between the new EOF and the maximum
1379 * possible file size.
1380 *
1381 * We have to free all the blocks to the bmbt maximum offset, even if
1382 * the page cache can't scale that far.
1383 */
1384 first_unmap_block = XFS_B_TO_FSB(mp, (xfs_ufsize_t)new_size);
1385 if (!xfs_verify_fileoff(mp, first_unmap_block)) {
1386 WARN_ON_ONCE(first_unmap_block > XFS_MAX_FILEOFF);
1387 return 0;
1388 }
1389
1390 unmap_len = XFS_MAX_FILEOFF - first_unmap_block + 1;
1391 while (unmap_len > 0) {
1392 ASSERT(tp->t_firstblock == NULLFSBLOCK);
1393 error = __xfs_bunmapi(tp, ip, first_unmap_block, &unmap_len,
1394 flags, XFS_ITRUNC_MAX_EXTENTS);
1395 if (error)
1396 goto out;
1397
1398 /* free the just unmapped extents */
1399 error = xfs_defer_finish(&tp);
1400 if (error)
1401 goto out;
1402 }
1403
1404 if (whichfork == XFS_DATA_FORK) {
1405 /* Remove all pending CoW reservations. */
1406 error = xfs_reflink_cancel_cow_blocks(ip, &tp,
1407 first_unmap_block, XFS_MAX_FILEOFF, true);
1408 if (error)
1409 goto out;
1410
1411 xfs_itruncate_clear_reflink_flags(ip);
1412 }
1413
1414 /*
1415 * Always re-log the inode so that our permanent transaction can keep
1416 * on rolling it forward in the log.
1417 */
1418 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1419
1420 trace_xfs_itruncate_extents_end(ip, new_size);
1421
1422out:
1423 *tpp = tp;
1424 return error;
1425}
1426
1427int
1428xfs_release(
1429 xfs_inode_t *ip)
1430{
1431 xfs_mount_t *mp = ip->i_mount;
1432 int error = 0;
1433
1434 if (!S_ISREG(VFS_I(ip)->i_mode) || (VFS_I(ip)->i_mode == 0))
1435 return 0;
1436
1437 /* If this is a read-only mount, don't do this (would generate I/O) */
1438 if (mp->m_flags & XFS_MOUNT_RDONLY)
1439 return 0;
1440
1441 if (!XFS_FORCED_SHUTDOWN(mp)) {
1442 int truncated;
1443
1444 /*
1445 * If we previously truncated this file and removed old data
1446 * in the process, we want to initiate "early" writeout on
1447 * the last close. This is an attempt to combat the notorious
1448 * NULL files problem which is particularly noticeable from a
1449 * truncate down, buffered (re-)write (delalloc), followed by
1450 * a crash. What we are effectively doing here is
1451 * significantly reducing the time window where we'd otherwise
1452 * be exposed to that problem.
1453 */
1454 truncated = xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED);
1455 if (truncated) {
1456 xfs_iflags_clear(ip, XFS_IDIRTY_RELEASE);
1457 if (ip->i_delayed_blks > 0) {
1458 error = filemap_flush(VFS_I(ip)->i_mapping);
1459 if (error)
1460 return error;
1461 }
1462 }
1463 }
1464
1465 if (VFS_I(ip)->i_nlink == 0)
1466 return 0;
1467
1468 /*
1469 * If we can't get the iolock just skip truncating the blocks past EOF
1470 * because we could deadlock with the mmap_lock otherwise. We'll get
1471 * another chance to drop them once the last reference to the inode is
1472 * dropped, so we'll never leak blocks permanently.
1473 */
1474 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL))
1475 return 0;
1476
1477 if (xfs_can_free_eofblocks(ip, false)) {
1478 /*
1479 * Check if the inode is being opened, written and closed
1480 * frequently and we have delayed allocation blocks outstanding
1481 * (e.g. streaming writes from the NFS server), truncating the
1482 * blocks past EOF will cause fragmentation to occur.
1483 *
1484 * In this case don't do the truncation, but we have to be
1485 * careful how we detect this case. Blocks beyond EOF show up as
1486 * i_delayed_blks even when the inode is clean, so we need to
1487 * truncate them away first before checking for a dirty release.
1488 * Hence on the first dirty close we will still remove the
1489 * speculative allocation, but after that we will leave it in
1490 * place.
1491 */
1492 if (xfs_iflags_test(ip, XFS_IDIRTY_RELEASE))
1493 goto out_unlock;
1494
1495 error = xfs_free_eofblocks(ip);
1496 if (error)
1497 goto out_unlock;
1498
1499 /* delalloc blocks after truncation means it really is dirty */
1500 if (ip->i_delayed_blks)
1501 xfs_iflags_set(ip, XFS_IDIRTY_RELEASE);
1502 }
1503
1504out_unlock:
1505 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1506 return error;
1507}
1508
1509/*
1510 * xfs_inactive_truncate
1511 *
1512 * Called to perform a truncate when an inode becomes unlinked.
1513 */
1514STATIC int
1515xfs_inactive_truncate(
1516 struct xfs_inode *ip)
1517{
1518 struct xfs_mount *mp = ip->i_mount;
1519 struct xfs_trans *tp;
1520 int error;
1521
1522 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate, 0, 0, 0, &tp);
1523 if (error) {
1524 ASSERT(XFS_FORCED_SHUTDOWN(mp));
1525 return error;
1526 }
1527 xfs_ilock(ip, XFS_ILOCK_EXCL);
1528 xfs_trans_ijoin(tp, ip, 0);
1529
1530 /*
1531 * Log the inode size first to prevent stale data exposure in the event
1532 * of a system crash before the truncate completes. See the related
1533 * comment in xfs_vn_setattr_size() for details.
1534 */
1535 ip->i_disk_size = 0;
1536 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1537
1538 error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK, 0);
1539 if (error)
1540 goto error_trans_cancel;
1541
1542 ASSERT(ip->i_df.if_nextents == 0);
1543
1544 error = xfs_trans_commit(tp);
1545 if (error)
1546 goto error_unlock;
1547
1548 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1549 return 0;
1550
1551error_trans_cancel:
1552 xfs_trans_cancel(tp);
1553error_unlock:
1554 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1555 return error;
1556}
1557
1558/*
1559 * xfs_inactive_ifree()
1560 *
1561 * Perform the inode free when an inode is unlinked.
1562 */
1563STATIC int
1564xfs_inactive_ifree(
1565 struct xfs_inode *ip)
1566{
1567 struct xfs_mount *mp = ip->i_mount;
1568 struct xfs_trans *tp;
1569 int error;
1570
1571 /*
1572 * We try to use a per-AG reservation for any block needed by the finobt
1573 * tree, but as the finobt feature predates the per-AG reservation
1574 * support a degraded file system might not have enough space for the
1575 * reservation at mount time. In that case try to dip into the reserved
1576 * pool and pray.
1577 *
1578 * Send a warning if the reservation does happen to fail, as the inode
1579 * now remains allocated and sits on the unlinked list until the fs is
1580 * repaired.
1581 */
1582 if (unlikely(mp->m_finobt_nores)) {
1583 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree,
1584 XFS_IFREE_SPACE_RES(mp), 0, XFS_TRANS_RESERVE,
1585 &tp);
1586 } else {
1587 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree, 0, 0, 0, &tp);
1588 }
1589 if (error) {
1590 if (error == -ENOSPC) {
1591 xfs_warn_ratelimited(mp,
1592 "Failed to remove inode(s) from unlinked list. "
1593 "Please free space, unmount and run xfs_repair.");
1594 } else {
1595 ASSERT(XFS_FORCED_SHUTDOWN(mp));
1596 }
1597 return error;
1598 }
1599
1600 /*
1601 * We do not hold the inode locked across the entire rolling transaction
1602 * here. We only need to hold it for the first transaction that
1603 * xfs_ifree() builds, which may mark the inode XFS_ISTALE if the
1604 * underlying cluster buffer is freed. Relogging an XFS_ISTALE inode
1605 * here breaks the relationship between cluster buffer invalidation and
1606 * stale inode invalidation on cluster buffer item journal commit
1607 * completion, and can result in leaving dirty stale inodes hanging
1608 * around in memory.
1609 *
1610 * We have no need for serialising this inode operation against other
1611 * operations - we freed the inode and hence reallocation is required
1612 * and that will serialise on reallocating the space the deferops need
1613 * to free. Hence we can unlock the inode on the first commit of
1614 * the transaction rather than roll it right through the deferops. This
1615 * avoids relogging the XFS_ISTALE inode.
1616 *
1617 * We check that xfs_ifree() hasn't grown an internal transaction roll
1618 * by asserting that the inode is still locked when it returns.
1619 */
1620 xfs_ilock(ip, XFS_ILOCK_EXCL);
1621 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1622
1623 error = xfs_ifree(tp, ip);
1624 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1625 if (error) {
1626 /*
1627 * If we fail to free the inode, shut down. The cancel
1628 * might do that, we need to make sure. Otherwise the
1629 * inode might be lost for a long time or forever.
1630 */
1631 if (!XFS_FORCED_SHUTDOWN(mp)) {
1632 xfs_notice(mp, "%s: xfs_ifree returned error %d",
1633 __func__, error);
1634 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1635 }
1636 xfs_trans_cancel(tp);
1637 return error;
1638 }
1639
1640 /*
1641 * Credit the quota account(s). The inode is gone.
1642 */
1643 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
1644
1645 /*
1646 * Just ignore errors at this point. There is nothing we can do except
1647 * to try to keep going. Make sure it's not a silent error.
1648 */
1649 error = xfs_trans_commit(tp);
1650 if (error)
1651 xfs_notice(mp, "%s: xfs_trans_commit returned error %d",
1652 __func__, error);
1653
1654 return 0;
1655}
1656
1657/*
1658 * xfs_inactive
1659 *
1660 * This is called when the vnode reference count for the vnode
1661 * goes to zero. If the file has been unlinked, then it must
1662 * now be truncated. Also, we clear all of the read-ahead state
1663 * kept for the inode here since the file is now closed.
1664 */
1665void
1666xfs_inactive(
1667 xfs_inode_t *ip)
1668{
1669 struct xfs_mount *mp;
1670 int error;
1671 int truncate = 0;
1672
1673 /*
1674 * If the inode is already free, then there can be nothing
1675 * to clean up here.
1676 */
1677 if (VFS_I(ip)->i_mode == 0) {
1678 ASSERT(ip->i_df.if_broot_bytes == 0);
1679 goto out;
1680 }
1681
1682 mp = ip->i_mount;
1683 ASSERT(!xfs_iflags_test(ip, XFS_IRECOVERY));
1684
1685 /* If this is a read-only mount, don't do this (would generate I/O) */
1686 if (mp->m_flags & XFS_MOUNT_RDONLY)
1687 goto out;
1688
1689 /* Metadata inodes require explicit resource cleanup. */
1690 if (xfs_is_metadata_inode(ip))
1691 goto out;
1692
1693 /* Try to clean out the cow blocks if there are any. */
1694 if (xfs_inode_has_cow_data(ip))
1695 xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, true);
1696
1697 if (VFS_I(ip)->i_nlink != 0) {
1698 /*
1699 * force is true because we are evicting an inode from the
1700 * cache. Post-eof blocks must be freed, lest we end up with
1701 * broken free space accounting.
1702 *
1703 * Note: don't bother with iolock here since lockdep complains
1704 * about acquiring it in reclaim context. We have the only
1705 * reference to the inode at this point anyways.
1706 */
1707 if (xfs_can_free_eofblocks(ip, true))
1708 xfs_free_eofblocks(ip);
1709
1710 goto out;
1711 }
1712
1713 if (S_ISREG(VFS_I(ip)->i_mode) &&
1714 (ip->i_disk_size != 0 || XFS_ISIZE(ip) != 0 ||
1715 ip->i_df.if_nextents > 0 || ip->i_delayed_blks > 0))
1716 truncate = 1;
1717
1718 error = xfs_qm_dqattach(ip);
1719 if (error)
1720 goto out;
1721
1722 if (S_ISLNK(VFS_I(ip)->i_mode))
1723 error = xfs_inactive_symlink(ip);
1724 else if (truncate)
1725 error = xfs_inactive_truncate(ip);
1726 if (error)
1727 goto out;
1728
1729 /*
1730 * If there are attributes associated with the file then blow them away
1731 * now. The code calls a routine that recursively deconstructs the
1732 * attribute fork. If also blows away the in-core attribute fork.
1733 */
1734 if (XFS_IFORK_Q(ip)) {
1735 error = xfs_attr_inactive(ip);
1736 if (error)
1737 goto out;
1738 }
1739
1740 ASSERT(!ip->i_afp);
1741 ASSERT(ip->i_forkoff == 0);
1742
1743 /*
1744 * Free the inode.
1745 */
1746 xfs_inactive_ifree(ip);
1747
1748out:
1749 /*
1750 * We're done making metadata updates for this inode, so we can release
1751 * the attached dquots.
1752 */
1753 xfs_qm_dqdetach(ip);
1754}
1755
1756/*
1757 * In-Core Unlinked List Lookups
1758 * =============================
1759 *
1760 * Every inode is supposed to be reachable from some other piece of metadata
1761 * with the exception of the root directory. Inodes with a connection to a
1762 * file descriptor but not linked from anywhere in the on-disk directory tree
1763 * are collectively known as unlinked inodes, though the filesystem itself
1764 * maintains links to these inodes so that on-disk metadata are consistent.
1765 *
1766 * XFS implements a per-AG on-disk hash table of unlinked inodes. The AGI
1767 * header contains a number of buckets that point to an inode, and each inode
1768 * record has a pointer to the next inode in the hash chain. This
1769 * singly-linked list causes scaling problems in the iunlink remove function
1770 * because we must walk that list to find the inode that points to the inode
1771 * being removed from the unlinked hash bucket list.
1772 *
1773 * What if we modelled the unlinked list as a collection of records capturing
1774 * "X.next_unlinked = Y" relations? If we indexed those records on Y, we'd
1775 * have a fast way to look up unlinked list predecessors, which avoids the
1776 * slow list walk. That's exactly what we do here (in-core) with a per-AG
1777 * rhashtable.
1778 *
1779 * Because this is a backref cache, we ignore operational failures since the
1780 * iunlink code can fall back to the slow bucket walk. The only errors that
1781 * should bubble out are for obviously incorrect situations.
1782 *
1783 * All users of the backref cache MUST hold the AGI buffer lock to serialize
1784 * access or have otherwise provided for concurrency control.
1785 */
1786
1787/* Capture a "X.next_unlinked = Y" relationship. */
1788struct xfs_iunlink {
1789 struct rhash_head iu_rhash_head;
1790 xfs_agino_t iu_agino; /* X */
1791 xfs_agino_t iu_next_unlinked; /* Y */
1792};
1793
1794/* Unlinked list predecessor lookup hashtable construction */
1795static int
1796xfs_iunlink_obj_cmpfn(
1797 struct rhashtable_compare_arg *arg,
1798 const void *obj)
1799{
1800 const xfs_agino_t *key = arg->key;
1801 const struct xfs_iunlink *iu = obj;
1802
1803 if (iu->iu_next_unlinked != *key)
1804 return 1;
1805 return 0;
1806}
1807
1808static const struct rhashtable_params xfs_iunlink_hash_params = {
1809 .min_size = XFS_AGI_UNLINKED_BUCKETS,
1810 .key_len = sizeof(xfs_agino_t),
1811 .key_offset = offsetof(struct xfs_iunlink,
1812 iu_next_unlinked),
1813 .head_offset = offsetof(struct xfs_iunlink, iu_rhash_head),
1814 .automatic_shrinking = true,
1815 .obj_cmpfn = xfs_iunlink_obj_cmpfn,
1816};
1817
1818/*
1819 * Return X, where X.next_unlinked == @agino. Returns NULLAGINO if no such
1820 * relation is found.
1821 */
1822static xfs_agino_t
1823xfs_iunlink_lookup_backref(
1824 struct xfs_perag *pag,
1825 xfs_agino_t agino)
1826{
1827 struct xfs_iunlink *iu;
1828
1829 iu = rhashtable_lookup_fast(&pag->pagi_unlinked_hash, &agino,
1830 xfs_iunlink_hash_params);
1831 return iu ? iu->iu_agino : NULLAGINO;
1832}
1833
1834/*
1835 * Take ownership of an iunlink cache entry and insert it into the hash table.
1836 * If successful, the entry will be owned by the cache; if not, it is freed.
1837 * Either way, the caller does not own @iu after this call.
1838 */
1839static int
1840xfs_iunlink_insert_backref(
1841 struct xfs_perag *pag,
1842 struct xfs_iunlink *iu)
1843{
1844 int error;
1845
1846 error = rhashtable_insert_fast(&pag->pagi_unlinked_hash,
1847 &iu->iu_rhash_head, xfs_iunlink_hash_params);
1848 /*
1849 * Fail loudly if there already was an entry because that's a sign of
1850 * corruption of in-memory data. Also fail loudly if we see an error
1851 * code we didn't anticipate from the rhashtable code. Currently we
1852 * only anticipate ENOMEM.
1853 */
1854 if (error) {
1855 WARN(error != -ENOMEM, "iunlink cache insert error %d", error);
1856 kmem_free(iu);
1857 }
1858 /*
1859 * Absorb any runtime errors that aren't a result of corruption because
1860 * this is a cache and we can always fall back to bucket list scanning.
1861 */
1862 if (error != 0 && error != -EEXIST)
1863 error = 0;
1864 return error;
1865}
1866
1867/* Remember that @prev_agino.next_unlinked = @this_agino. */
1868static int
1869xfs_iunlink_add_backref(
1870 struct xfs_perag *pag,
1871 xfs_agino_t prev_agino,
1872 xfs_agino_t this_agino)
1873{
1874 struct xfs_iunlink *iu;
1875
1876 if (XFS_TEST_ERROR(false, pag->pag_mount, XFS_ERRTAG_IUNLINK_FALLBACK))
1877 return 0;
1878
1879 iu = kmem_zalloc(sizeof(*iu), KM_NOFS);
1880 iu->iu_agino = prev_agino;
1881 iu->iu_next_unlinked = this_agino;
1882
1883 return xfs_iunlink_insert_backref(pag, iu);
1884}
1885
1886/*
1887 * Replace X.next_unlinked = @agino with X.next_unlinked = @next_unlinked.
1888 * If @next_unlinked is NULLAGINO, we drop the backref and exit. If there
1889 * wasn't any such entry then we don't bother.
1890 */
1891static int
1892xfs_iunlink_change_backref(
1893 struct xfs_perag *pag,
1894 xfs_agino_t agino,
1895 xfs_agino_t next_unlinked)
1896{
1897 struct xfs_iunlink *iu;
1898 int error;
1899
1900 /* Look up the old entry; if there wasn't one then exit. */
1901 iu = rhashtable_lookup_fast(&pag->pagi_unlinked_hash, &agino,
1902 xfs_iunlink_hash_params);
1903 if (!iu)
1904 return 0;
1905
1906 /*
1907 * Remove the entry. This shouldn't ever return an error, but if we
1908 * couldn't remove the old entry we don't want to add it again to the
1909 * hash table, and if the entry disappeared on us then someone's
1910 * violated the locking rules and we need to fail loudly. Either way
1911 * we cannot remove the inode because internal state is or would have
1912 * been corrupt.
1913 */
1914 error = rhashtable_remove_fast(&pag->pagi_unlinked_hash,
1915 &iu->iu_rhash_head, xfs_iunlink_hash_params);
1916 if (error)
1917 return error;
1918
1919 /* If there is no new next entry just free our item and return. */
1920 if (next_unlinked == NULLAGINO) {
1921 kmem_free(iu);
1922 return 0;
1923 }
1924
1925 /* Update the entry and re-add it to the hash table. */
1926 iu->iu_next_unlinked = next_unlinked;
1927 return xfs_iunlink_insert_backref(pag, iu);
1928}
1929
1930/* Set up the in-core predecessor structures. */
1931int
1932xfs_iunlink_init(
1933 struct xfs_perag *pag)
1934{
1935 return rhashtable_init(&pag->pagi_unlinked_hash,
1936 &xfs_iunlink_hash_params);
1937}
1938
1939/* Free the in-core predecessor structures. */
1940static void
1941xfs_iunlink_free_item(
1942 void *ptr,
1943 void *arg)
1944{
1945 struct xfs_iunlink *iu = ptr;
1946 bool *freed_anything = arg;
1947
1948 *freed_anything = true;
1949 kmem_free(iu);
1950}
1951
1952void
1953xfs_iunlink_destroy(
1954 struct xfs_perag *pag)
1955{
1956 bool freed_anything = false;
1957
1958 rhashtable_free_and_destroy(&pag->pagi_unlinked_hash,
1959 xfs_iunlink_free_item, &freed_anything);
1960
1961 ASSERT(freed_anything == false || XFS_FORCED_SHUTDOWN(pag->pag_mount));
1962}
1963
1964/*
1965 * Point the AGI unlinked bucket at an inode and log the results. The caller
1966 * is responsible for validating the old value.
1967 */
1968STATIC int
1969xfs_iunlink_update_bucket(
1970 struct xfs_trans *tp,
1971 struct xfs_perag *pag,
1972 struct xfs_buf *agibp,
1973 unsigned int bucket_index,
1974 xfs_agino_t new_agino)
1975{
1976 struct xfs_agi *agi = agibp->b_addr;
1977 xfs_agino_t old_value;
1978 int offset;
1979
1980 ASSERT(xfs_verify_agino_or_null(tp->t_mountp, pag->pag_agno, new_agino));
1981
1982 old_value = be32_to_cpu(agi->agi_unlinked[bucket_index]);
1983 trace_xfs_iunlink_update_bucket(tp->t_mountp, pag->pag_agno, bucket_index,
1984 old_value, new_agino);
1985
1986 /*
1987 * We should never find the head of the list already set to the value
1988 * passed in because either we're adding or removing ourselves from the
1989 * head of the list.
1990 */
1991 if (old_value == new_agino) {
1992 xfs_buf_mark_corrupt(agibp);
1993 return -EFSCORRUPTED;
1994 }
1995
1996 agi->agi_unlinked[bucket_index] = cpu_to_be32(new_agino);
1997 offset = offsetof(struct xfs_agi, agi_unlinked) +
1998 (sizeof(xfs_agino_t) * bucket_index);
1999 xfs_trans_log_buf(tp, agibp, offset, offset + sizeof(xfs_agino_t) - 1);
2000 return 0;
2001}
2002
2003/* Set an on-disk inode's next_unlinked pointer. */
2004STATIC void
2005xfs_iunlink_update_dinode(
2006 struct xfs_trans *tp,
2007 struct xfs_perag *pag,
2008 xfs_agino_t agino,
2009 struct xfs_buf *ibp,
2010 struct xfs_dinode *dip,
2011 struct xfs_imap *imap,
2012 xfs_agino_t next_agino)
2013{
2014 struct xfs_mount *mp = tp->t_mountp;
2015 int offset;
2016
2017 ASSERT(xfs_verify_agino_or_null(mp, pag->pag_agno, next_agino));
2018
2019 trace_xfs_iunlink_update_dinode(mp, pag->pag_agno, agino,
2020 be32_to_cpu(dip->di_next_unlinked), next_agino);
2021
2022 dip->di_next_unlinked = cpu_to_be32(next_agino);
2023 offset = imap->im_boffset +
2024 offsetof(struct xfs_dinode, di_next_unlinked);
2025
2026 /* need to recalc the inode CRC if appropriate */
2027 xfs_dinode_calc_crc(mp, dip);
2028 xfs_trans_inode_buf(tp, ibp);
2029 xfs_trans_log_buf(tp, ibp, offset, offset + sizeof(xfs_agino_t) - 1);
2030}
2031
2032/* Set an in-core inode's unlinked pointer and return the old value. */
2033STATIC int
2034xfs_iunlink_update_inode(
2035 struct xfs_trans *tp,
2036 struct xfs_inode *ip,
2037 struct xfs_perag *pag,
2038 xfs_agino_t next_agino,
2039 xfs_agino_t *old_next_agino)
2040{
2041 struct xfs_mount *mp = tp->t_mountp;
2042 struct xfs_dinode *dip;
2043 struct xfs_buf *ibp;
2044 xfs_agino_t old_value;
2045 int error;
2046
2047 ASSERT(xfs_verify_agino_or_null(mp, pag->pag_agno, next_agino));
2048
2049 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &ibp);
2050 if (error)
2051 return error;
2052 dip = xfs_buf_offset(ibp, ip->i_imap.im_boffset);
2053
2054 /* Make sure the old pointer isn't garbage. */
2055 old_value = be32_to_cpu(dip->di_next_unlinked);
2056 if (!xfs_verify_agino_or_null(mp, pag->pag_agno, old_value)) {
2057 xfs_inode_verifier_error(ip, -EFSCORRUPTED, __func__, dip,
2058 sizeof(*dip), __this_address);
2059 error = -EFSCORRUPTED;
2060 goto out;
2061 }
2062
2063 /*
2064 * Since we're updating a linked list, we should never find that the
2065 * current pointer is the same as the new value, unless we're
2066 * terminating the list.
2067 */
2068 *old_next_agino = old_value;
2069 if (old_value == next_agino) {
2070 if (next_agino != NULLAGINO) {
2071 xfs_inode_verifier_error(ip, -EFSCORRUPTED, __func__,
2072 dip, sizeof(*dip), __this_address);
2073 error = -EFSCORRUPTED;
2074 }
2075 goto out;
2076 }
2077
2078 /* Ok, update the new pointer. */
2079 xfs_iunlink_update_dinode(tp, pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
2080 ibp, dip, &ip->i_imap, next_agino);
2081 return 0;
2082out:
2083 xfs_trans_brelse(tp, ibp);
2084 return error;
2085}
2086
2087/*
2088 * This is called when the inode's link count has gone to 0 or we are creating
2089 * a tmpfile via O_TMPFILE. The inode @ip must have nlink == 0.
2090 *
2091 * We place the on-disk inode on a list in the AGI. It will be pulled from this
2092 * list when the inode is freed.
2093 */
2094STATIC int
2095xfs_iunlink(
2096 struct xfs_trans *tp,
2097 struct xfs_inode *ip)
2098{
2099 struct xfs_mount *mp = tp->t_mountp;
2100 struct xfs_perag *pag;
2101 struct xfs_agi *agi;
2102 struct xfs_buf *agibp;
2103 xfs_agino_t next_agino;
2104 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2105 short bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
2106 int error;
2107
2108 ASSERT(VFS_I(ip)->i_nlink == 0);
2109 ASSERT(VFS_I(ip)->i_mode != 0);
2110 trace_xfs_iunlink(ip);
2111
2112 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
2113
2114 /* Get the agi buffer first. It ensures lock ordering on the list. */
2115 error = xfs_read_agi(mp, tp, pag->pag_agno, &agibp);
2116 if (error)
2117 goto out;
2118 agi = agibp->b_addr;
2119
2120 /*
2121 * Get the index into the agi hash table for the list this inode will
2122 * go on. Make sure the pointer isn't garbage and that this inode
2123 * isn't already on the list.
2124 */
2125 next_agino = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2126 if (next_agino == agino ||
2127 !xfs_verify_agino_or_null(mp, pag->pag_agno, next_agino)) {
2128 xfs_buf_mark_corrupt(agibp);
2129 error = -EFSCORRUPTED;
2130 goto out;
2131 }
2132
2133 if (next_agino != NULLAGINO) {
2134 xfs_agino_t old_agino;
2135
2136 /*
2137 * There is already another inode in the bucket, so point this
2138 * inode to the current head of the list.
2139 */
2140 error = xfs_iunlink_update_inode(tp, ip, pag, next_agino,
2141 &old_agino);
2142 if (error)
2143 goto out;
2144 ASSERT(old_agino == NULLAGINO);
2145
2146 /*
2147 * agino has been unlinked, add a backref from the next inode
2148 * back to agino.
2149 */
2150 error = xfs_iunlink_add_backref(pag, agino, next_agino);
2151 if (error)
2152 goto out;
2153 }
2154
2155 /* Point the head of the list to point to this inode. */
2156 error = xfs_iunlink_update_bucket(tp, pag, agibp, bucket_index, agino);
2157out:
2158 xfs_perag_put(pag);
2159 return error;
2160}
2161
2162/* Return the imap, dinode pointer, and buffer for an inode. */
2163STATIC int
2164xfs_iunlink_map_ino(
2165 struct xfs_trans *tp,
2166 xfs_agnumber_t agno,
2167 xfs_agino_t agino,
2168 struct xfs_imap *imap,
2169 struct xfs_dinode **dipp,
2170 struct xfs_buf **bpp)
2171{
2172 struct xfs_mount *mp = tp->t_mountp;
2173 int error;
2174
2175 imap->im_blkno = 0;
2176 error = xfs_imap(mp, tp, XFS_AGINO_TO_INO(mp, agno, agino), imap, 0);
2177 if (error) {
2178 xfs_warn(mp, "%s: xfs_imap returned error %d.",
2179 __func__, error);
2180 return error;
2181 }
2182
2183 error = xfs_imap_to_bp(mp, tp, imap, bpp);
2184 if (error) {
2185 xfs_warn(mp, "%s: xfs_imap_to_bp returned error %d.",
2186 __func__, error);
2187 return error;
2188 }
2189
2190 *dipp = xfs_buf_offset(*bpp, imap->im_boffset);
2191 return 0;
2192}
2193
2194/*
2195 * Walk the unlinked chain from @head_agino until we find the inode that
2196 * points to @target_agino. Return the inode number, map, dinode pointer,
2197 * and inode cluster buffer of that inode as @agino, @imap, @dipp, and @bpp.
2198 *
2199 * @tp, @pag, @head_agino, and @target_agino are input parameters.
2200 * @agino, @imap, @dipp, and @bpp are all output parameters.
2201 *
2202 * Do not call this function if @target_agino is the head of the list.
2203 */
2204STATIC int
2205xfs_iunlink_map_prev(
2206 struct xfs_trans *tp,
2207 struct xfs_perag *pag,
2208 xfs_agino_t head_agino,
2209 xfs_agino_t target_agino,
2210 xfs_agino_t *agino,
2211 struct xfs_imap *imap,
2212 struct xfs_dinode **dipp,
2213 struct xfs_buf **bpp)
2214{
2215 struct xfs_mount *mp = tp->t_mountp;
2216 xfs_agino_t next_agino;
2217 int error;
2218
2219 ASSERT(head_agino != target_agino);
2220 *bpp = NULL;
2221
2222 /* See if our backref cache can find it faster. */
2223 *agino = xfs_iunlink_lookup_backref(pag, target_agino);
2224 if (*agino != NULLAGINO) {
2225 error = xfs_iunlink_map_ino(tp, pag->pag_agno, *agino, imap,
2226 dipp, bpp);
2227 if (error)
2228 return error;
2229
2230 if (be32_to_cpu((*dipp)->di_next_unlinked) == target_agino)
2231 return 0;
2232
2233 /*
2234 * If we get here the cache contents were corrupt, so drop the
2235 * buffer and fall back to walking the bucket list.
2236 */
2237 xfs_trans_brelse(tp, *bpp);
2238 *bpp = NULL;
2239 WARN_ON_ONCE(1);
2240 }
2241
2242 trace_xfs_iunlink_map_prev_fallback(mp, pag->pag_agno);
2243
2244 /* Otherwise, walk the entire bucket until we find it. */
2245 next_agino = head_agino;
2246 while (next_agino != target_agino) {
2247 xfs_agino_t unlinked_agino;
2248
2249 if (*bpp)
2250 xfs_trans_brelse(tp, *bpp);
2251
2252 *agino = next_agino;
2253 error = xfs_iunlink_map_ino(tp, pag->pag_agno, next_agino, imap,
2254 dipp, bpp);
2255 if (error)
2256 return error;
2257
2258 unlinked_agino = be32_to_cpu((*dipp)->di_next_unlinked);
2259 /*
2260 * Make sure this pointer is valid and isn't an obvious
2261 * infinite loop.
2262 */
2263 if (!xfs_verify_agino(mp, pag->pag_agno, unlinked_agino) ||
2264 next_agino == unlinked_agino) {
2265 XFS_CORRUPTION_ERROR(__func__,
2266 XFS_ERRLEVEL_LOW, mp,
2267 *dipp, sizeof(**dipp));
2268 error = -EFSCORRUPTED;
2269 return error;
2270 }
2271 next_agino = unlinked_agino;
2272 }
2273
2274 return 0;
2275}
2276
2277/*
2278 * Pull the on-disk inode from the AGI unlinked list.
2279 */
2280STATIC int
2281xfs_iunlink_remove(
2282 struct xfs_trans *tp,
2283 struct xfs_perag *pag,
2284 struct xfs_inode *ip)
2285{
2286 struct xfs_mount *mp = tp->t_mountp;
2287 struct xfs_agi *agi;
2288 struct xfs_buf *agibp;
2289 struct xfs_buf *last_ibp;
2290 struct xfs_dinode *last_dip = NULL;
2291 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2292 xfs_agino_t next_agino;
2293 xfs_agino_t head_agino;
2294 short bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
2295 int error;
2296
2297 trace_xfs_iunlink_remove(ip);
2298
2299 /* Get the agi buffer first. It ensures lock ordering on the list. */
2300 error = xfs_read_agi(mp, tp, pag->pag_agno, &agibp);
2301 if (error)
2302 return error;
2303 agi = agibp->b_addr;
2304
2305 /*
2306 * Get the index into the agi hash table for the list this inode will
2307 * go on. Make sure the head pointer isn't garbage.
2308 */
2309 head_agino = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2310 if (!xfs_verify_agino(mp, pag->pag_agno, head_agino)) {
2311 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
2312 agi, sizeof(*agi));
2313 return -EFSCORRUPTED;
2314 }
2315
2316 /*
2317 * Set our inode's next_unlinked pointer to NULL and then return
2318 * the old pointer value so that we can update whatever was previous
2319 * to us in the list to point to whatever was next in the list.
2320 */
2321 error = xfs_iunlink_update_inode(tp, ip, pag, NULLAGINO, &next_agino);
2322 if (error)
2323 return error;
2324
2325 /*
2326 * If there was a backref pointing from the next inode back to this
2327 * one, remove it because we've removed this inode from the list.
2328 *
2329 * Later, if this inode was in the middle of the list we'll update
2330 * this inode's backref to point from the next inode.
2331 */
2332 if (next_agino != NULLAGINO) {
2333 error = xfs_iunlink_change_backref(pag, next_agino, NULLAGINO);
2334 if (error)
2335 return error;
2336 }
2337
2338 if (head_agino != agino) {
2339 struct xfs_imap imap;
2340 xfs_agino_t prev_agino;
2341
2342 /* We need to search the list for the inode being freed. */
2343 error = xfs_iunlink_map_prev(tp, pag, head_agino, agino,
2344 &prev_agino, &imap, &last_dip, &last_ibp);
2345 if (error)
2346 return error;
2347
2348 /* Point the previous inode on the list to the next inode. */
2349 xfs_iunlink_update_dinode(tp, pag, prev_agino, last_ibp,
2350 last_dip, &imap, next_agino);
2351
2352 /*
2353 * Now we deal with the backref for this inode. If this inode
2354 * pointed at a real inode, change the backref that pointed to
2355 * us to point to our old next. If this inode was the end of
2356 * the list, delete the backref that pointed to us. Note that
2357 * change_backref takes care of deleting the backref if
2358 * next_agino is NULLAGINO.
2359 */
2360 return xfs_iunlink_change_backref(agibp->b_pag, agino,
2361 next_agino);
2362 }
2363
2364 /* Point the head of the list to the next unlinked inode. */
2365 return xfs_iunlink_update_bucket(tp, pag, agibp, bucket_index,
2366 next_agino);
2367}
2368
2369/*
2370 * Look up the inode number specified and if it is not already marked XFS_ISTALE
2371 * mark it stale. We should only find clean inodes in this lookup that aren't
2372 * already stale.
2373 */
2374static void
2375xfs_ifree_mark_inode_stale(
2376 struct xfs_perag *pag,
2377 struct xfs_inode *free_ip,
2378 xfs_ino_t inum)
2379{
2380 struct xfs_mount *mp = pag->pag_mount;
2381 struct xfs_inode_log_item *iip;
2382 struct xfs_inode *ip;
2383
2384retry:
2385 rcu_read_lock();
2386 ip = radix_tree_lookup(&pag->pag_ici_root, XFS_INO_TO_AGINO(mp, inum));
2387
2388 /* Inode not in memory, nothing to do */
2389 if (!ip) {
2390 rcu_read_unlock();
2391 return;
2392 }
2393
2394 /*
2395 * because this is an RCU protected lookup, we could find a recently
2396 * freed or even reallocated inode during the lookup. We need to check
2397 * under the i_flags_lock for a valid inode here. Skip it if it is not
2398 * valid, the wrong inode or stale.
2399 */
2400 spin_lock(&ip->i_flags_lock);
2401 if (ip->i_ino != inum || __xfs_iflags_test(ip, XFS_ISTALE))
2402 goto out_iflags_unlock;
2403
2404 /*
2405 * Don't try to lock/unlock the current inode, but we _cannot_ skip the
2406 * other inodes that we did not find in the list attached to the buffer
2407 * and are not already marked stale. If we can't lock it, back off and
2408 * retry.
2409 */
2410 if (ip != free_ip) {
2411 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL)) {
2412 spin_unlock(&ip->i_flags_lock);
2413 rcu_read_unlock();
2414 delay(1);
2415 goto retry;
2416 }
2417 }
2418 ip->i_flags |= XFS_ISTALE;
2419
2420 /*
2421 * If the inode is flushing, it is already attached to the buffer. All
2422 * we needed to do here is mark the inode stale so buffer IO completion
2423 * will remove it from the AIL.
2424 */
2425 iip = ip->i_itemp;
2426 if (__xfs_iflags_test(ip, XFS_IFLUSHING)) {
2427 ASSERT(!list_empty(&iip->ili_item.li_bio_list));
2428 ASSERT(iip->ili_last_fields);
2429 goto out_iunlock;
2430 }
2431
2432 /*
2433 * Inodes not attached to the buffer can be released immediately.
2434 * Everything else has to go through xfs_iflush_abort() on journal
2435 * commit as the flock synchronises removal of the inode from the
2436 * cluster buffer against inode reclaim.
2437 */
2438 if (!iip || list_empty(&iip->ili_item.li_bio_list))
2439 goto out_iunlock;
2440
2441 __xfs_iflags_set(ip, XFS_IFLUSHING);
2442 spin_unlock(&ip->i_flags_lock);
2443 rcu_read_unlock();
2444
2445 /* we have a dirty inode in memory that has not yet been flushed. */
2446 spin_lock(&iip->ili_lock);
2447 iip->ili_last_fields = iip->ili_fields;
2448 iip->ili_fields = 0;
2449 iip->ili_fsync_fields = 0;
2450 spin_unlock(&iip->ili_lock);
2451 ASSERT(iip->ili_last_fields);
2452
2453 if (ip != free_ip)
2454 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2455 return;
2456
2457out_iunlock:
2458 if (ip != free_ip)
2459 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2460out_iflags_unlock:
2461 spin_unlock(&ip->i_flags_lock);
2462 rcu_read_unlock();
2463}
2464
2465/*
2466 * A big issue when freeing the inode cluster is that we _cannot_ skip any
2467 * inodes that are in memory - they all must be marked stale and attached to
2468 * the cluster buffer.
2469 */
2470static int
2471xfs_ifree_cluster(
2472 struct xfs_trans *tp,
2473 struct xfs_perag *pag,
2474 struct xfs_inode *free_ip,
2475 struct xfs_icluster *xic)
2476{
2477 struct xfs_mount *mp = free_ip->i_mount;
2478 struct xfs_ino_geometry *igeo = M_IGEO(mp);
2479 struct xfs_buf *bp;
2480 xfs_daddr_t blkno;
2481 xfs_ino_t inum = xic->first_ino;
2482 int nbufs;
2483 int i, j;
2484 int ioffset;
2485 int error;
2486
2487 nbufs = igeo->ialloc_blks / igeo->blocks_per_cluster;
2488
2489 for (j = 0; j < nbufs; j++, inum += igeo->inodes_per_cluster) {
2490 /*
2491 * The allocation bitmap tells us which inodes of the chunk were
2492 * physically allocated. Skip the cluster if an inode falls into
2493 * a sparse region.
2494 */
2495 ioffset = inum - xic->first_ino;
2496 if ((xic->alloc & XFS_INOBT_MASK(ioffset)) == 0) {
2497 ASSERT(ioffset % igeo->inodes_per_cluster == 0);
2498 continue;
2499 }
2500
2501 blkno = XFS_AGB_TO_DADDR(mp, XFS_INO_TO_AGNO(mp, inum),
2502 XFS_INO_TO_AGBNO(mp, inum));
2503
2504 /*
2505 * We obtain and lock the backing buffer first in the process
2506 * here to ensure dirty inodes attached to the buffer remain in
2507 * the flushing state while we mark them stale.
2508 *
2509 * If we scan the in-memory inodes first, then buffer IO can
2510 * complete before we get a lock on it, and hence we may fail
2511 * to mark all the active inodes on the buffer stale.
2512 */
2513 error = xfs_trans_get_buf(tp, mp->m_ddev_targp, blkno,
2514 mp->m_bsize * igeo->blocks_per_cluster,
2515 XBF_UNMAPPED, &bp);
2516 if (error)
2517 return error;
2518
2519 /*
2520 * This buffer may not have been correctly initialised as we
2521 * didn't read it from disk. That's not important because we are
2522 * only using to mark the buffer as stale in the log, and to
2523 * attach stale cached inodes on it. That means it will never be
2524 * dispatched for IO. If it is, we want to know about it, and we
2525 * want it to fail. We can acheive this by adding a write
2526 * verifier to the buffer.
2527 */
2528 bp->b_ops = &xfs_inode_buf_ops;
2529
2530 /*
2531 * Now we need to set all the cached clean inodes as XFS_ISTALE,
2532 * too. This requires lookups, and will skip inodes that we've
2533 * already marked XFS_ISTALE.
2534 */
2535 for (i = 0; i < igeo->inodes_per_cluster; i++)
2536 xfs_ifree_mark_inode_stale(pag, free_ip, inum + i);
2537
2538 xfs_trans_stale_inode_buf(tp, bp);
2539 xfs_trans_binval(tp, bp);
2540 }
2541 return 0;
2542}
2543
2544/*
2545 * This is called to return an inode to the inode free list.
2546 * The inode should already be truncated to 0 length and have
2547 * no pages associated with it. This routine also assumes that
2548 * the inode is already a part of the transaction.
2549 *
2550 * The on-disk copy of the inode will have been added to the list
2551 * of unlinked inodes in the AGI. We need to remove the inode from
2552 * that list atomically with respect to freeing it here.
2553 */
2554int
2555xfs_ifree(
2556 struct xfs_trans *tp,
2557 struct xfs_inode *ip)
2558{
2559 struct xfs_mount *mp = ip->i_mount;
2560 struct xfs_perag *pag;
2561 struct xfs_icluster xic = { 0 };
2562 struct xfs_inode_log_item *iip = ip->i_itemp;
2563 int error;
2564
2565 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
2566 ASSERT(VFS_I(ip)->i_nlink == 0);
2567 ASSERT(ip->i_df.if_nextents == 0);
2568 ASSERT(ip->i_disk_size == 0 || !S_ISREG(VFS_I(ip)->i_mode));
2569 ASSERT(ip->i_nblocks == 0);
2570
2571 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
2572
2573 /*
2574 * Pull the on-disk inode from the AGI unlinked list.
2575 */
2576 error = xfs_iunlink_remove(tp, pag, ip);
2577 if (error)
2578 goto out;
2579
2580 error = xfs_difree(tp, pag, ip->i_ino, &xic);
2581 if (error)
2582 goto out;
2583
2584 /*
2585 * Free any local-format data sitting around before we reset the
2586 * data fork to extents format. Note that the attr fork data has
2587 * already been freed by xfs_attr_inactive.
2588 */
2589 if (ip->i_df.if_format == XFS_DINODE_FMT_LOCAL) {
2590 kmem_free(ip->i_df.if_u1.if_data);
2591 ip->i_df.if_u1.if_data = NULL;
2592 ip->i_df.if_bytes = 0;
2593 }
2594
2595 VFS_I(ip)->i_mode = 0; /* mark incore inode as free */
2596 ip->i_diflags = 0;
2597 ip->i_diflags2 = mp->m_ino_geo.new_diflags2;
2598 ip->i_forkoff = 0; /* mark the attr fork not in use */
2599 ip->i_df.if_format = XFS_DINODE_FMT_EXTENTS;
2600 if (xfs_iflags_test(ip, XFS_IPRESERVE_DM_FIELDS))
2601 xfs_iflags_clear(ip, XFS_IPRESERVE_DM_FIELDS);
2602
2603 /* Don't attempt to replay owner changes for a deleted inode */
2604 spin_lock(&iip->ili_lock);
2605 iip->ili_fields &= ~(XFS_ILOG_AOWNER | XFS_ILOG_DOWNER);
2606 spin_unlock(&iip->ili_lock);
2607
2608 /*
2609 * Bump the generation count so no one will be confused
2610 * by reincarnations of this inode.
2611 */
2612 VFS_I(ip)->i_generation++;
2613 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2614
2615 if (xic.deleted)
2616 error = xfs_ifree_cluster(tp, pag, ip, &xic);
2617out:
2618 xfs_perag_put(pag);
2619 return error;
2620}
2621
2622/*
2623 * This is called to unpin an inode. The caller must have the inode locked
2624 * in at least shared mode so that the buffer cannot be subsequently pinned
2625 * once someone is waiting for it to be unpinned.
2626 */
2627static void
2628xfs_iunpin(
2629 struct xfs_inode *ip)
2630{
2631 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
2632
2633 trace_xfs_inode_unpin_nowait(ip, _RET_IP_);
2634
2635 /* Give the log a push to start the unpinning I/O */
2636 xfs_log_force_seq(ip->i_mount, ip->i_itemp->ili_commit_seq, 0, NULL);
2637
2638}
2639
2640static void
2641__xfs_iunpin_wait(
2642 struct xfs_inode *ip)
2643{
2644 wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_IPINNED_BIT);
2645 DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IPINNED_BIT);
2646
2647 xfs_iunpin(ip);
2648
2649 do {
2650 prepare_to_wait(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
2651 if (xfs_ipincount(ip))
2652 io_schedule();
2653 } while (xfs_ipincount(ip));
2654 finish_wait(wq, &wait.wq_entry);
2655}
2656
2657void
2658xfs_iunpin_wait(
2659 struct xfs_inode *ip)
2660{
2661 if (xfs_ipincount(ip))
2662 __xfs_iunpin_wait(ip);
2663}
2664
2665/*
2666 * Removing an inode from the namespace involves removing the directory entry
2667 * and dropping the link count on the inode. Removing the directory entry can
2668 * result in locking an AGF (directory blocks were freed) and removing a link
2669 * count can result in placing the inode on an unlinked list which results in
2670 * locking an AGI.
2671 *
2672 * The big problem here is that we have an ordering constraint on AGF and AGI
2673 * locking - inode allocation locks the AGI, then can allocate a new extent for
2674 * new inodes, locking the AGF after the AGI. Similarly, freeing the inode
2675 * removes the inode from the unlinked list, requiring that we lock the AGI
2676 * first, and then freeing the inode can result in an inode chunk being freed
2677 * and hence freeing disk space requiring that we lock an AGF.
2678 *
2679 * Hence the ordering that is imposed by other parts of the code is AGI before
2680 * AGF. This means we cannot remove the directory entry before we drop the inode
2681 * reference count and put it on the unlinked list as this results in a lock
2682 * order of AGF then AGI, and this can deadlock against inode allocation and
2683 * freeing. Therefore we must drop the link counts before we remove the
2684 * directory entry.
2685 *
2686 * This is still safe from a transactional point of view - it is not until we
2687 * get to xfs_defer_finish() that we have the possibility of multiple
2688 * transactions in this operation. Hence as long as we remove the directory
2689 * entry and drop the link count in the first transaction of the remove
2690 * operation, there are no transactional constraints on the ordering here.
2691 */
2692int
2693xfs_remove(
2694 xfs_inode_t *dp,
2695 struct xfs_name *name,
2696 xfs_inode_t *ip)
2697{
2698 xfs_mount_t *mp = dp->i_mount;
2699 xfs_trans_t *tp = NULL;
2700 int is_dir = S_ISDIR(VFS_I(ip)->i_mode);
2701 int error = 0;
2702 uint resblks;
2703
2704 trace_xfs_remove(dp, name);
2705
2706 if (XFS_FORCED_SHUTDOWN(mp))
2707 return -EIO;
2708
2709 error = xfs_qm_dqattach(dp);
2710 if (error)
2711 goto std_return;
2712
2713 error = xfs_qm_dqattach(ip);
2714 if (error)
2715 goto std_return;
2716
2717 /*
2718 * We try to get the real space reservation first,
2719 * allowing for directory btree deletion(s) implying
2720 * possible bmap insert(s). If we can't get the space
2721 * reservation then we use 0 instead, and avoid the bmap
2722 * btree insert(s) in the directory code by, if the bmap
2723 * insert tries to happen, instead trimming the LAST
2724 * block from the directory.
2725 */
2726 resblks = XFS_REMOVE_SPACE_RES(mp);
2727 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_remove, resblks, 0, 0, &tp);
2728 if (error == -ENOSPC) {
2729 resblks = 0;
2730 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_remove, 0, 0, 0,
2731 &tp);
2732 }
2733 if (error) {
2734 ASSERT(error != -ENOSPC);
2735 goto std_return;
2736 }
2737
2738 xfs_lock_two_inodes(dp, XFS_ILOCK_EXCL, ip, XFS_ILOCK_EXCL);
2739
2740 xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
2741 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
2742
2743 /*
2744 * If we're removing a directory perform some additional validation.
2745 */
2746 if (is_dir) {
2747 ASSERT(VFS_I(ip)->i_nlink >= 2);
2748 if (VFS_I(ip)->i_nlink != 2) {
2749 error = -ENOTEMPTY;
2750 goto out_trans_cancel;
2751 }
2752 if (!xfs_dir_isempty(ip)) {
2753 error = -ENOTEMPTY;
2754 goto out_trans_cancel;
2755 }
2756
2757 /* Drop the link from ip's "..". */
2758 error = xfs_droplink(tp, dp);
2759 if (error)
2760 goto out_trans_cancel;
2761
2762 /* Drop the "." link from ip to self. */
2763 error = xfs_droplink(tp, ip);
2764 if (error)
2765 goto out_trans_cancel;
2766
2767 /*
2768 * Point the unlinked child directory's ".." entry to the root
2769 * directory to eliminate back-references to inodes that may
2770 * get freed before the child directory is closed. If the fs
2771 * gets shrunk, this can lead to dirent inode validation errors.
2772 */
2773 if (dp->i_ino != tp->t_mountp->m_sb.sb_rootino) {
2774 error = xfs_dir_replace(tp, ip, &xfs_name_dotdot,
2775 tp->t_mountp->m_sb.sb_rootino, 0);
2776 if (error)
2777 return error;
2778 }
2779 } else {
2780 /*
2781 * When removing a non-directory we need to log the parent
2782 * inode here. For a directory this is done implicitly
2783 * by the xfs_droplink call for the ".." entry.
2784 */
2785 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
2786 }
2787 xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2788
2789 /* Drop the link from dp to ip. */
2790 error = xfs_droplink(tp, ip);
2791 if (error)
2792 goto out_trans_cancel;
2793
2794 error = xfs_dir_removename(tp, dp, name, ip->i_ino, resblks);
2795 if (error) {
2796 ASSERT(error != -ENOENT);
2797 goto out_trans_cancel;
2798 }
2799
2800 /*
2801 * If this is a synchronous mount, make sure that the
2802 * remove transaction goes to disk before returning to
2803 * the user.
2804 */
2805 if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
2806 xfs_trans_set_sync(tp);
2807
2808 error = xfs_trans_commit(tp);
2809 if (error)
2810 goto std_return;
2811
2812 if (is_dir && xfs_inode_is_filestream(ip))
2813 xfs_filestream_deassociate(ip);
2814
2815 return 0;
2816
2817 out_trans_cancel:
2818 xfs_trans_cancel(tp);
2819 std_return:
2820 return error;
2821}
2822
2823/*
2824 * Enter all inodes for a rename transaction into a sorted array.
2825 */
2826#define __XFS_SORT_INODES 5
2827STATIC void
2828xfs_sort_for_rename(
2829 struct xfs_inode *dp1, /* in: old (source) directory inode */
2830 struct xfs_inode *dp2, /* in: new (target) directory inode */
2831 struct xfs_inode *ip1, /* in: inode of old entry */
2832 struct xfs_inode *ip2, /* in: inode of new entry */
2833 struct xfs_inode *wip, /* in: whiteout inode */
2834 struct xfs_inode **i_tab,/* out: sorted array of inodes */
2835 int *num_inodes) /* in/out: inodes in array */
2836{
2837 int i, j;
2838
2839 ASSERT(*num_inodes == __XFS_SORT_INODES);
2840 memset(i_tab, 0, *num_inodes * sizeof(struct xfs_inode *));
2841
2842 /*
2843 * i_tab contains a list of pointers to inodes. We initialize
2844 * the table here & we'll sort it. We will then use it to
2845 * order the acquisition of the inode locks.
2846 *
2847 * Note that the table may contain duplicates. e.g., dp1 == dp2.
2848 */
2849 i = 0;
2850 i_tab[i++] = dp1;
2851 i_tab[i++] = dp2;
2852 i_tab[i++] = ip1;
2853 if (ip2)
2854 i_tab[i++] = ip2;
2855 if (wip)
2856 i_tab[i++] = wip;
2857 *num_inodes = i;
2858
2859 /*
2860 * Sort the elements via bubble sort. (Remember, there are at
2861 * most 5 elements to sort, so this is adequate.)
2862 */
2863 for (i = 0; i < *num_inodes; i++) {
2864 for (j = 1; j < *num_inodes; j++) {
2865 if (i_tab[j]->i_ino < i_tab[j-1]->i_ino) {
2866 struct xfs_inode *temp = i_tab[j];
2867 i_tab[j] = i_tab[j-1];
2868 i_tab[j-1] = temp;
2869 }
2870 }
2871 }
2872}
2873
2874static int
2875xfs_finish_rename(
2876 struct xfs_trans *tp)
2877{
2878 /*
2879 * If this is a synchronous mount, make sure that the rename transaction
2880 * goes to disk before returning to the user.
2881 */
2882 if (tp->t_mountp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
2883 xfs_trans_set_sync(tp);
2884
2885 return xfs_trans_commit(tp);
2886}
2887
2888/*
2889 * xfs_cross_rename()
2890 *
2891 * responsible for handling RENAME_EXCHANGE flag in renameat2() syscall
2892 */
2893STATIC int
2894xfs_cross_rename(
2895 struct xfs_trans *tp,
2896 struct xfs_inode *dp1,
2897 struct xfs_name *name1,
2898 struct xfs_inode *ip1,
2899 struct xfs_inode *dp2,
2900 struct xfs_name *name2,
2901 struct xfs_inode *ip2,
2902 int spaceres)
2903{
2904 int error = 0;
2905 int ip1_flags = 0;
2906 int ip2_flags = 0;
2907 int dp2_flags = 0;
2908
2909 /* Swap inode number for dirent in first parent */
2910 error = xfs_dir_replace(tp, dp1, name1, ip2->i_ino, spaceres);
2911 if (error)
2912 goto out_trans_abort;
2913
2914 /* Swap inode number for dirent in second parent */
2915 error = xfs_dir_replace(tp, dp2, name2, ip1->i_ino, spaceres);
2916 if (error)
2917 goto out_trans_abort;
2918
2919 /*
2920 * If we're renaming one or more directories across different parents,
2921 * update the respective ".." entries (and link counts) to match the new
2922 * parents.
2923 */
2924 if (dp1 != dp2) {
2925 dp2_flags = XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2926
2927 if (S_ISDIR(VFS_I(ip2)->i_mode)) {
2928 error = xfs_dir_replace(tp, ip2, &xfs_name_dotdot,
2929 dp1->i_ino, spaceres);
2930 if (error)
2931 goto out_trans_abort;
2932
2933 /* transfer ip2 ".." reference to dp1 */
2934 if (!S_ISDIR(VFS_I(ip1)->i_mode)) {
2935 error = xfs_droplink(tp, dp2);
2936 if (error)
2937 goto out_trans_abort;
2938 xfs_bumplink(tp, dp1);
2939 }
2940
2941 /*
2942 * Although ip1 isn't changed here, userspace needs
2943 * to be warned about the change, so that applications
2944 * relying on it (like backup ones), will properly
2945 * notify the change
2946 */
2947 ip1_flags |= XFS_ICHGTIME_CHG;
2948 ip2_flags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2949 }
2950
2951 if (S_ISDIR(VFS_I(ip1)->i_mode)) {
2952 error = xfs_dir_replace(tp, ip1, &xfs_name_dotdot,
2953 dp2->i_ino, spaceres);
2954 if (error)
2955 goto out_trans_abort;
2956
2957 /* transfer ip1 ".." reference to dp2 */
2958 if (!S_ISDIR(VFS_I(ip2)->i_mode)) {
2959 error = xfs_droplink(tp, dp1);
2960 if (error)
2961 goto out_trans_abort;
2962 xfs_bumplink(tp, dp2);
2963 }
2964
2965 /*
2966 * Although ip2 isn't changed here, userspace needs
2967 * to be warned about the change, so that applications
2968 * relying on it (like backup ones), will properly
2969 * notify the change
2970 */
2971 ip1_flags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2972 ip2_flags |= XFS_ICHGTIME_CHG;
2973 }
2974 }
2975
2976 if (ip1_flags) {
2977 xfs_trans_ichgtime(tp, ip1, ip1_flags);
2978 xfs_trans_log_inode(tp, ip1, XFS_ILOG_CORE);
2979 }
2980 if (ip2_flags) {
2981 xfs_trans_ichgtime(tp, ip2, ip2_flags);
2982 xfs_trans_log_inode(tp, ip2, XFS_ILOG_CORE);
2983 }
2984 if (dp2_flags) {
2985 xfs_trans_ichgtime(tp, dp2, dp2_flags);
2986 xfs_trans_log_inode(tp, dp2, XFS_ILOG_CORE);
2987 }
2988 xfs_trans_ichgtime(tp, dp1, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2989 xfs_trans_log_inode(tp, dp1, XFS_ILOG_CORE);
2990 return xfs_finish_rename(tp);
2991
2992out_trans_abort:
2993 xfs_trans_cancel(tp);
2994 return error;
2995}
2996
2997/*
2998 * xfs_rename_alloc_whiteout()
2999 *
3000 * Return a referenced, unlinked, unlocked inode that can be used as a
3001 * whiteout in a rename transaction. We use a tmpfile inode here so that if we
3002 * crash between allocating the inode and linking it into the rename transaction
3003 * recovery will free the inode and we won't leak it.
3004 */
3005static int
3006xfs_rename_alloc_whiteout(
3007 struct user_namespace *mnt_userns,
3008 struct xfs_inode *dp,
3009 struct xfs_inode **wip)
3010{
3011 struct xfs_inode *tmpfile;
3012 int error;
3013
3014 error = xfs_create_tmpfile(mnt_userns, dp, S_IFCHR | WHITEOUT_MODE,
3015 &tmpfile);
3016 if (error)
3017 return error;
3018
3019 /*
3020 * Prepare the tmpfile inode as if it were created through the VFS.
3021 * Complete the inode setup and flag it as linkable. nlink is already
3022 * zero, so we can skip the drop_nlink.
3023 */
3024 xfs_setup_iops(tmpfile);
3025 xfs_finish_inode_setup(tmpfile);
3026 VFS_I(tmpfile)->i_state |= I_LINKABLE;
3027
3028 *wip = tmpfile;
3029 return 0;
3030}
3031
3032/*
3033 * xfs_rename
3034 */
3035int
3036xfs_rename(
3037 struct user_namespace *mnt_userns,
3038 struct xfs_inode *src_dp,
3039 struct xfs_name *src_name,
3040 struct xfs_inode *src_ip,
3041 struct xfs_inode *target_dp,
3042 struct xfs_name *target_name,
3043 struct xfs_inode *target_ip,
3044 unsigned int flags)
3045{
3046 struct xfs_mount *mp = src_dp->i_mount;
3047 struct xfs_trans *tp;
3048 struct xfs_inode *wip = NULL; /* whiteout inode */
3049 struct xfs_inode *inodes[__XFS_SORT_INODES];
3050 int i;
3051 int num_inodes = __XFS_SORT_INODES;
3052 bool new_parent = (src_dp != target_dp);
3053 bool src_is_directory = S_ISDIR(VFS_I(src_ip)->i_mode);
3054 int spaceres;
3055 int error;
3056
3057 trace_xfs_rename(src_dp, target_dp, src_name, target_name);
3058
3059 if ((flags & RENAME_EXCHANGE) && !target_ip)
3060 return -EINVAL;
3061
3062 /*
3063 * If we are doing a whiteout operation, allocate the whiteout inode
3064 * we will be placing at the target and ensure the type is set
3065 * appropriately.
3066 */
3067 if (flags & RENAME_WHITEOUT) {
3068 ASSERT(!(flags & (RENAME_NOREPLACE | RENAME_EXCHANGE)));
3069 error = xfs_rename_alloc_whiteout(mnt_userns, target_dp, &wip);
3070 if (error)
3071 return error;
3072
3073 /* setup target dirent info as whiteout */
3074 src_name->type = XFS_DIR3_FT_CHRDEV;
3075 }
3076
3077 xfs_sort_for_rename(src_dp, target_dp, src_ip, target_ip, wip,
3078 inodes, &num_inodes);
3079
3080 spaceres = XFS_RENAME_SPACE_RES(mp, target_name->len);
3081 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, spaceres, 0, 0, &tp);
3082 if (error == -ENOSPC) {
3083 spaceres = 0;
3084 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, 0, 0, 0,
3085 &tp);
3086 }
3087 if (error)
3088 goto out_release_wip;
3089
3090 /*
3091 * Attach the dquots to the inodes
3092 */
3093 error = xfs_qm_vop_rename_dqattach(inodes);
3094 if (error)
3095 goto out_trans_cancel;
3096
3097 /*
3098 * Lock all the participating inodes. Depending upon whether
3099 * the target_name exists in the target directory, and
3100 * whether the target directory is the same as the source
3101 * directory, we can lock from 2 to 4 inodes.
3102 */
3103 xfs_lock_inodes(inodes, num_inodes, XFS_ILOCK_EXCL);
3104
3105 /*
3106 * Join all the inodes to the transaction. From this point on,
3107 * we can rely on either trans_commit or trans_cancel to unlock
3108 * them.
3109 */
3110 xfs_trans_ijoin(tp, src_dp, XFS_ILOCK_EXCL);
3111 if (new_parent)
3112 xfs_trans_ijoin(tp, target_dp, XFS_ILOCK_EXCL);
3113 xfs_trans_ijoin(tp, src_ip, XFS_ILOCK_EXCL);
3114 if (target_ip)
3115 xfs_trans_ijoin(tp, target_ip, XFS_ILOCK_EXCL);
3116 if (wip)
3117 xfs_trans_ijoin(tp, wip, XFS_ILOCK_EXCL);
3118
3119 /*
3120 * If we are using project inheritance, we only allow renames
3121 * into our tree when the project IDs are the same; else the
3122 * tree quota mechanism would be circumvented.
3123 */
3124 if (unlikely((target_dp->i_diflags & XFS_DIFLAG_PROJINHERIT) &&
3125 target_dp->i_projid != src_ip->i_projid)) {
3126 error = -EXDEV;
3127 goto out_trans_cancel;
3128 }
3129
3130 /* RENAME_EXCHANGE is unique from here on. */
3131 if (flags & RENAME_EXCHANGE)
3132 return xfs_cross_rename(tp, src_dp, src_name, src_ip,
3133 target_dp, target_name, target_ip,
3134 spaceres);
3135
3136 /*
3137 * Check for expected errors before we dirty the transaction
3138 * so we can return an error without a transaction abort.
3139 *
3140 * Extent count overflow check:
3141 *
3142 * From the perspective of src_dp, a rename operation is essentially a
3143 * directory entry remove operation. Hence the only place where we check
3144 * for extent count overflow for src_dp is in
3145 * xfs_bmap_del_extent_real(). xfs_bmap_del_extent_real() returns
3146 * -ENOSPC when it detects a possible extent count overflow and in
3147 * response, the higher layers of directory handling code do the
3148 * following:
3149 * 1. Data/Free blocks: XFS lets these blocks linger until a
3150 * future remove operation removes them.
3151 * 2. Dabtree blocks: XFS swaps the blocks with the last block in the
3152 * Leaf space and unmaps the last block.
3153 *
3154 * For target_dp, there are two cases depending on whether the
3155 * destination directory entry exists or not.
3156 *
3157 * When destination directory entry does not exist (i.e. target_ip ==
3158 * NULL), extent count overflow check is performed only when transaction
3159 * has a non-zero sized space reservation associated with it. With a
3160 * zero-sized space reservation, XFS allows a rename operation to
3161 * continue only when the directory has sufficient free space in its
3162 * data/leaf/free space blocks to hold the new entry.
3163 *
3164 * When destination directory entry exists (i.e. target_ip != NULL), all
3165 * we need to do is change the inode number associated with the already
3166 * existing entry. Hence there is no need to perform an extent count
3167 * overflow check.
3168 */
3169 if (target_ip == NULL) {
3170 /*
3171 * If there's no space reservation, check the entry will
3172 * fit before actually inserting it.
3173 */
3174 if (!spaceres) {
3175 error = xfs_dir_canenter(tp, target_dp, target_name);
3176 if (error)
3177 goto out_trans_cancel;
3178 } else {
3179 error = xfs_iext_count_may_overflow(target_dp,
3180 XFS_DATA_FORK,
3181 XFS_IEXT_DIR_MANIP_CNT(mp));
3182 if (error)
3183 goto out_trans_cancel;
3184 }
3185 } else {
3186 /*
3187 * If target exists and it's a directory, check that whether
3188 * it can be destroyed.
3189 */
3190 if (S_ISDIR(VFS_I(target_ip)->i_mode) &&
3191 (!xfs_dir_isempty(target_ip) ||
3192 (VFS_I(target_ip)->i_nlink > 2))) {
3193 error = -EEXIST;
3194 goto out_trans_cancel;
3195 }
3196 }
3197
3198 /*
3199 * Lock the AGI buffers we need to handle bumping the nlink of the
3200 * whiteout inode off the unlinked list and to handle dropping the
3201 * nlink of the target inode. Per locking order rules, do this in
3202 * increasing AG order and before directory block allocation tries to
3203 * grab AGFs because we grab AGIs before AGFs.
3204 *
3205 * The (vfs) caller must ensure that if src is a directory then
3206 * target_ip is either null or an empty directory.
3207 */
3208 for (i = 0; i < num_inodes && inodes[i] != NULL; i++) {
3209 if (inodes[i] == wip ||
3210 (inodes[i] == target_ip &&
3211 (VFS_I(target_ip)->i_nlink == 1 || src_is_directory))) {
3212 struct xfs_buf *bp;
3213 xfs_agnumber_t agno;
3214
3215 agno = XFS_INO_TO_AGNO(mp, inodes[i]->i_ino);
3216 error = xfs_read_agi(mp, tp, agno, &bp);
3217 if (error)
3218 goto out_trans_cancel;
3219 }
3220 }
3221
3222 /*
3223 * Directory entry creation below may acquire the AGF. Remove
3224 * the whiteout from the unlinked list first to preserve correct
3225 * AGI/AGF locking order. This dirties the transaction so failures
3226 * after this point will abort and log recovery will clean up the
3227 * mess.
3228 *
3229 * For whiteouts, we need to bump the link count on the whiteout
3230 * inode. After this point, we have a real link, clear the tmpfile
3231 * state flag from the inode so it doesn't accidentally get misused
3232 * in future.
3233 */
3234 if (wip) {
3235 struct xfs_perag *pag;
3236
3237 ASSERT(VFS_I(wip)->i_nlink == 0);
3238
3239 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, wip->i_ino));
3240 error = xfs_iunlink_remove(tp, pag, wip);
3241 xfs_perag_put(pag);
3242 if (error)
3243 goto out_trans_cancel;
3244
3245 xfs_bumplink(tp, wip);
3246 VFS_I(wip)->i_state &= ~I_LINKABLE;
3247 }
3248
3249 /*
3250 * Set up the target.
3251 */
3252 if (target_ip == NULL) {
3253 /*
3254 * If target does not exist and the rename crosses
3255 * directories, adjust the target directory link count
3256 * to account for the ".." reference from the new entry.
3257 */
3258 error = xfs_dir_createname(tp, target_dp, target_name,
3259 src_ip->i_ino, spaceres);
3260 if (error)
3261 goto out_trans_cancel;
3262
3263 xfs_trans_ichgtime(tp, target_dp,
3264 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3265
3266 if (new_parent && src_is_directory) {
3267 xfs_bumplink(tp, target_dp);
3268 }
3269 } else { /* target_ip != NULL */
3270 /*
3271 * Link the source inode under the target name.
3272 * If the source inode is a directory and we are moving
3273 * it across directories, its ".." entry will be
3274 * inconsistent until we replace that down below.
3275 *
3276 * In case there is already an entry with the same
3277 * name at the destination directory, remove it first.
3278 */
3279 error = xfs_dir_replace(tp, target_dp, target_name,
3280 src_ip->i_ino, spaceres);
3281 if (error)
3282 goto out_trans_cancel;
3283
3284 xfs_trans_ichgtime(tp, target_dp,
3285 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3286
3287 /*
3288 * Decrement the link count on the target since the target
3289 * dir no longer points to it.
3290 */
3291 error = xfs_droplink(tp, target_ip);
3292 if (error)
3293 goto out_trans_cancel;
3294
3295 if (src_is_directory) {
3296 /*
3297 * Drop the link from the old "." entry.
3298 */
3299 error = xfs_droplink(tp, target_ip);
3300 if (error)
3301 goto out_trans_cancel;
3302 }
3303 } /* target_ip != NULL */
3304
3305 /*
3306 * Remove the source.
3307 */
3308 if (new_parent && src_is_directory) {
3309 /*
3310 * Rewrite the ".." entry to point to the new
3311 * directory.
3312 */
3313 error = xfs_dir_replace(tp, src_ip, &xfs_name_dotdot,
3314 target_dp->i_ino, spaceres);
3315 ASSERT(error != -EEXIST);
3316 if (error)
3317 goto out_trans_cancel;
3318 }
3319
3320 /*
3321 * We always want to hit the ctime on the source inode.
3322 *
3323 * This isn't strictly required by the standards since the source
3324 * inode isn't really being changed, but old unix file systems did
3325 * it and some incremental backup programs won't work without it.
3326 */
3327 xfs_trans_ichgtime(tp, src_ip, XFS_ICHGTIME_CHG);
3328 xfs_trans_log_inode(tp, src_ip, XFS_ILOG_CORE);
3329
3330 /*
3331 * Adjust the link count on src_dp. This is necessary when
3332 * renaming a directory, either within one parent when
3333 * the target existed, or across two parent directories.
3334 */
3335 if (src_is_directory && (new_parent || target_ip != NULL)) {
3336
3337 /*
3338 * Decrement link count on src_directory since the
3339 * entry that's moved no longer points to it.
3340 */
3341 error = xfs_droplink(tp, src_dp);
3342 if (error)
3343 goto out_trans_cancel;
3344 }
3345
3346 /*
3347 * For whiteouts, we only need to update the source dirent with the
3348 * inode number of the whiteout inode rather than removing it
3349 * altogether.
3350 */
3351 if (wip) {
3352 error = xfs_dir_replace(tp, src_dp, src_name, wip->i_ino,
3353 spaceres);
3354 } else {
3355 /*
3356 * NOTE: We don't need to check for extent count overflow here
3357 * because the dir remove name code will leave the dir block in
3358 * place if the extent count would overflow.
3359 */
3360 error = xfs_dir_removename(tp, src_dp, src_name, src_ip->i_ino,
3361 spaceres);
3362 }
3363
3364 if (error)
3365 goto out_trans_cancel;
3366
3367 xfs_trans_ichgtime(tp, src_dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3368 xfs_trans_log_inode(tp, src_dp, XFS_ILOG_CORE);
3369 if (new_parent)
3370 xfs_trans_log_inode(tp, target_dp, XFS_ILOG_CORE);
3371
3372 error = xfs_finish_rename(tp);
3373 if (wip)
3374 xfs_irele(wip);
3375 return error;
3376
3377out_trans_cancel:
3378 xfs_trans_cancel(tp);
3379out_release_wip:
3380 if (wip)
3381 xfs_irele(wip);
3382 return error;
3383}
3384
3385static int
3386xfs_iflush(
3387 struct xfs_inode *ip,
3388 struct xfs_buf *bp)
3389{
3390 struct xfs_inode_log_item *iip = ip->i_itemp;
3391 struct xfs_dinode *dip;
3392 struct xfs_mount *mp = ip->i_mount;
3393 int error;
3394
3395 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
3396 ASSERT(xfs_iflags_test(ip, XFS_IFLUSHING));
3397 ASSERT(ip->i_df.if_format != XFS_DINODE_FMT_BTREE ||
3398 ip->i_df.if_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
3399 ASSERT(iip->ili_item.li_buf == bp);
3400
3401 dip = xfs_buf_offset(bp, ip->i_imap.im_boffset);
3402
3403 /*
3404 * We don't flush the inode if any of the following checks fail, but we
3405 * do still update the log item and attach to the backing buffer as if
3406 * the flush happened. This is a formality to facilitate predictable
3407 * error handling as the caller will shutdown and fail the buffer.
3408 */
3409 error = -EFSCORRUPTED;
3410 if (XFS_TEST_ERROR(dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC),
3411 mp, XFS_ERRTAG_IFLUSH_1)) {
3412 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3413 "%s: Bad inode %Lu magic number 0x%x, ptr "PTR_FMT,
3414 __func__, ip->i_ino, be16_to_cpu(dip->di_magic), dip);
3415 goto flush_out;
3416 }
3417 if (S_ISREG(VFS_I(ip)->i_mode)) {
3418 if (XFS_TEST_ERROR(
3419 ip->i_df.if_format != XFS_DINODE_FMT_EXTENTS &&
3420 ip->i_df.if_format != XFS_DINODE_FMT_BTREE,
3421 mp, XFS_ERRTAG_IFLUSH_3)) {
3422 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3423 "%s: Bad regular inode %Lu, ptr "PTR_FMT,
3424 __func__, ip->i_ino, ip);
3425 goto flush_out;
3426 }
3427 } else if (S_ISDIR(VFS_I(ip)->i_mode)) {
3428 if (XFS_TEST_ERROR(
3429 ip->i_df.if_format != XFS_DINODE_FMT_EXTENTS &&
3430 ip->i_df.if_format != XFS_DINODE_FMT_BTREE &&
3431 ip->i_df.if_format != XFS_DINODE_FMT_LOCAL,
3432 mp, XFS_ERRTAG_IFLUSH_4)) {
3433 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3434 "%s: Bad directory inode %Lu, ptr "PTR_FMT,
3435 __func__, ip->i_ino, ip);
3436 goto flush_out;
3437 }
3438 }
3439 if (XFS_TEST_ERROR(ip->i_df.if_nextents + xfs_ifork_nextents(ip->i_afp) >
3440 ip->i_nblocks, mp, XFS_ERRTAG_IFLUSH_5)) {
3441 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3442 "%s: detected corrupt incore inode %Lu, "
3443 "total extents = %d, nblocks = %Ld, ptr "PTR_FMT,
3444 __func__, ip->i_ino,
3445 ip->i_df.if_nextents + xfs_ifork_nextents(ip->i_afp),
3446 ip->i_nblocks, ip);
3447 goto flush_out;
3448 }
3449 if (XFS_TEST_ERROR(ip->i_forkoff > mp->m_sb.sb_inodesize,
3450 mp, XFS_ERRTAG_IFLUSH_6)) {
3451 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3452 "%s: bad inode %Lu, forkoff 0x%x, ptr "PTR_FMT,
3453 __func__, ip->i_ino, ip->i_forkoff, ip);
3454 goto flush_out;
3455 }
3456
3457 /*
3458 * Inode item log recovery for v2 inodes are dependent on the flushiter
3459 * count for correct sequencing. We bump the flush iteration count so
3460 * we can detect flushes which postdate a log record during recovery.
3461 * This is redundant as we now log every change and hence this can't
3462 * happen but we need to still do it to ensure backwards compatibility
3463 * with old kernels that predate logging all inode changes.
3464 */
3465 if (!xfs_sb_version_has_v3inode(&mp->m_sb))
3466 ip->i_flushiter++;
3467
3468 /*
3469 * If there are inline format data / attr forks attached to this inode,
3470 * make sure they are not corrupt.
3471 */
3472 if (ip->i_df.if_format == XFS_DINODE_FMT_LOCAL &&
3473 xfs_ifork_verify_local_data(ip))
3474 goto flush_out;
3475 if (ip->i_afp && ip->i_afp->if_format == XFS_DINODE_FMT_LOCAL &&
3476 xfs_ifork_verify_local_attr(ip))
3477 goto flush_out;
3478
3479 /*
3480 * Copy the dirty parts of the inode into the on-disk inode. We always
3481 * copy out the core of the inode, because if the inode is dirty at all
3482 * the core must be.
3483 */
3484 xfs_inode_to_disk(ip, dip, iip->ili_item.li_lsn);
3485
3486 /* Wrap, we never let the log put out DI_MAX_FLUSH */
3487 if (!xfs_sb_version_has_v3inode(&mp->m_sb)) {
3488 if (ip->i_flushiter == DI_MAX_FLUSH)
3489 ip->i_flushiter = 0;
3490 }
3491
3492 xfs_iflush_fork(ip, dip, iip, XFS_DATA_FORK);
3493 if (XFS_IFORK_Q(ip))
3494 xfs_iflush_fork(ip, dip, iip, XFS_ATTR_FORK);
3495
3496 /*
3497 * We've recorded everything logged in the inode, so we'd like to clear
3498 * the ili_fields bits so we don't log and flush things unnecessarily.
3499 * However, we can't stop logging all this information until the data
3500 * we've copied into the disk buffer is written to disk. If we did we
3501 * might overwrite the copy of the inode in the log with all the data
3502 * after re-logging only part of it, and in the face of a crash we
3503 * wouldn't have all the data we need to recover.
3504 *
3505 * What we do is move the bits to the ili_last_fields field. When
3506 * logging the inode, these bits are moved back to the ili_fields field.
3507 * In the xfs_buf_inode_iodone() routine we clear ili_last_fields, since
3508 * we know that the information those bits represent is permanently on
3509 * disk. As long as the flush completes before the inode is logged
3510 * again, then both ili_fields and ili_last_fields will be cleared.
3511 */
3512 error = 0;
3513flush_out:
3514 spin_lock(&iip->ili_lock);
3515 iip->ili_last_fields = iip->ili_fields;
3516 iip->ili_fields = 0;
3517 iip->ili_fsync_fields = 0;
3518 spin_unlock(&iip->ili_lock);
3519
3520 /*
3521 * Store the current LSN of the inode so that we can tell whether the
3522 * item has moved in the AIL from xfs_buf_inode_iodone().
3523 */
3524 xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
3525 &iip->ili_item.li_lsn);
3526
3527 /* generate the checksum. */
3528 xfs_dinode_calc_crc(mp, dip);
3529 return error;
3530}
3531
3532/*
3533 * Non-blocking flush of dirty inode metadata into the backing buffer.
3534 *
3535 * The caller must have a reference to the inode and hold the cluster buffer
3536 * locked. The function will walk across all the inodes on the cluster buffer it
3537 * can find and lock without blocking, and flush them to the cluster buffer.
3538 *
3539 * On successful flushing of at least one inode, the caller must write out the
3540 * buffer and release it. If no inodes are flushed, -EAGAIN will be returned and
3541 * the caller needs to release the buffer. On failure, the filesystem will be
3542 * shut down, the buffer will have been unlocked and released, and EFSCORRUPTED
3543 * will be returned.
3544 */
3545int
3546xfs_iflush_cluster(
3547 struct xfs_buf *bp)
3548{
3549 struct xfs_mount *mp = bp->b_mount;
3550 struct xfs_log_item *lip, *n;
3551 struct xfs_inode *ip;
3552 struct xfs_inode_log_item *iip;
3553 int clcount = 0;
3554 int error = 0;
3555
3556 /*
3557 * We must use the safe variant here as on shutdown xfs_iflush_abort()
3558 * can remove itself from the list.
3559 */
3560 list_for_each_entry_safe(lip, n, &bp->b_li_list, li_bio_list) {
3561 iip = (struct xfs_inode_log_item *)lip;
3562 ip = iip->ili_inode;
3563
3564 /*
3565 * Quick and dirty check to avoid locks if possible.
3566 */
3567 if (__xfs_iflags_test(ip, XFS_IRECLAIM | XFS_IFLUSHING))
3568 continue;
3569 if (xfs_ipincount(ip))
3570 continue;
3571
3572 /*
3573 * The inode is still attached to the buffer, which means it is
3574 * dirty but reclaim might try to grab it. Check carefully for
3575 * that, and grab the ilock while still holding the i_flags_lock
3576 * to guarantee reclaim will not be able to reclaim this inode
3577 * once we drop the i_flags_lock.
3578 */
3579 spin_lock(&ip->i_flags_lock);
3580 ASSERT(!__xfs_iflags_test(ip, XFS_ISTALE));
3581 if (__xfs_iflags_test(ip, XFS_IRECLAIM | XFS_IFLUSHING)) {
3582 spin_unlock(&ip->i_flags_lock);
3583 continue;
3584 }
3585
3586 /*
3587 * ILOCK will pin the inode against reclaim and prevent
3588 * concurrent transactions modifying the inode while we are
3589 * flushing the inode. If we get the lock, set the flushing
3590 * state before we drop the i_flags_lock.
3591 */
3592 if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
3593 spin_unlock(&ip->i_flags_lock);
3594 continue;
3595 }
3596 __xfs_iflags_set(ip, XFS_IFLUSHING);
3597 spin_unlock(&ip->i_flags_lock);
3598
3599 /*
3600 * Abort flushing this inode if we are shut down because the
3601 * inode may not currently be in the AIL. This can occur when
3602 * log I/O failure unpins the inode without inserting into the
3603 * AIL, leaving a dirty/unpinned inode attached to the buffer
3604 * that otherwise looks like it should be flushed.
3605 */
3606 if (XFS_FORCED_SHUTDOWN(mp)) {
3607 xfs_iunpin_wait(ip);
3608 xfs_iflush_abort(ip);
3609 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3610 error = -EIO;
3611 continue;
3612 }
3613
3614 /* don't block waiting on a log force to unpin dirty inodes */
3615 if (xfs_ipincount(ip)) {
3616 xfs_iflags_clear(ip, XFS_IFLUSHING);
3617 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3618 continue;
3619 }
3620
3621 if (!xfs_inode_clean(ip))
3622 error = xfs_iflush(ip, bp);
3623 else
3624 xfs_iflags_clear(ip, XFS_IFLUSHING);
3625 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3626 if (error)
3627 break;
3628 clcount++;
3629 }
3630
3631 if (error) {
3632 bp->b_flags |= XBF_ASYNC;
3633 xfs_buf_ioend_fail(bp);
3634 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
3635 return error;
3636 }
3637
3638 if (!clcount)
3639 return -EAGAIN;
3640
3641 XFS_STATS_INC(mp, xs_icluster_flushcnt);
3642 XFS_STATS_ADD(mp, xs_icluster_flushinode, clcount);
3643 return 0;
3644
3645}
3646
3647/* Release an inode. */
3648void
3649xfs_irele(
3650 struct xfs_inode *ip)
3651{
3652 trace_xfs_irele(ip, _RET_IP_);
3653 iput(VFS_I(ip));
3654}
3655
3656/*
3657 * Ensure all commited transactions touching the inode are written to the log.
3658 */
3659int
3660xfs_log_force_inode(
3661 struct xfs_inode *ip)
3662{
3663 xfs_csn_t seq = 0;
3664
3665 xfs_ilock(ip, XFS_ILOCK_SHARED);
3666 if (xfs_ipincount(ip))
3667 seq = ip->i_itemp->ili_commit_seq;
3668 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3669
3670 if (!seq)
3671 return 0;
3672 return xfs_log_force_seq(ip->i_mount, seq, XFS_LOG_SYNC, NULL);
3673}
3674
3675/*
3676 * Grab the exclusive iolock for a data copy from src to dest, making sure to
3677 * abide vfs locking order (lowest pointer value goes first) and breaking the
3678 * layout leases before proceeding. The loop is needed because we cannot call
3679 * the blocking break_layout() with the iolocks held, and therefore have to
3680 * back out both locks.
3681 */
3682static int
3683xfs_iolock_two_inodes_and_break_layout(
3684 struct inode *src,
3685 struct inode *dest)
3686{
3687 int error;
3688
3689 if (src > dest)
3690 swap(src, dest);
3691
3692retry:
3693 /* Wait to break both inodes' layouts before we start locking. */
3694 error = break_layout(src, true);
3695 if (error)
3696 return error;
3697 if (src != dest) {
3698 error = break_layout(dest, true);
3699 if (error)
3700 return error;
3701 }
3702
3703 /* Lock one inode and make sure nobody got in and leased it. */
3704 inode_lock(src);
3705 error = break_layout(src, false);
3706 if (error) {
3707 inode_unlock(src);
3708 if (error == -EWOULDBLOCK)
3709 goto retry;
3710 return error;
3711 }
3712
3713 if (src == dest)
3714 return 0;
3715
3716 /* Lock the other inode and make sure nobody got in and leased it. */
3717 inode_lock_nested(dest, I_MUTEX_NONDIR2);
3718 error = break_layout(dest, false);
3719 if (error) {
3720 inode_unlock(src);
3721 inode_unlock(dest);
3722 if (error == -EWOULDBLOCK)
3723 goto retry;
3724 return error;
3725 }
3726
3727 return 0;
3728}
3729
3730/*
3731 * Lock two inodes so that userspace cannot initiate I/O via file syscalls or
3732 * mmap activity.
3733 */
3734int
3735xfs_ilock2_io_mmap(
3736 struct xfs_inode *ip1,
3737 struct xfs_inode *ip2)
3738{
3739 int ret;
3740
3741 ret = xfs_iolock_two_inodes_and_break_layout(VFS_I(ip1), VFS_I(ip2));
3742 if (ret)
3743 return ret;
3744 if (ip1 == ip2)
3745 xfs_ilock(ip1, XFS_MMAPLOCK_EXCL);
3746 else
3747 xfs_lock_two_inodes(ip1, XFS_MMAPLOCK_EXCL,
3748 ip2, XFS_MMAPLOCK_EXCL);
3749 return 0;
3750}
3751
3752/* Unlock both inodes to allow IO and mmap activity. */
3753void
3754xfs_iunlock2_io_mmap(
3755 struct xfs_inode *ip1,
3756 struct xfs_inode *ip2)
3757{
3758 bool same_inode = (ip1 == ip2);
3759
3760 xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
3761 if (!same_inode)
3762 xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
3763 inode_unlock(VFS_I(ip2));
3764 if (!same_inode)
3765 inode_unlock(VFS_I(ip1));
3766}