Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
4 * All Rights Reserved.
5 */
6#include <linux/iversion.h>
7
8#include "xfs.h"
9#include "xfs_fs.h"
10#include "xfs_shared.h"
11#include "xfs_format.h"
12#include "xfs_log_format.h"
13#include "xfs_trans_resv.h"
14#include "xfs_mount.h"
15#include "xfs_defer.h"
16#include "xfs_inode.h"
17#include "xfs_dir2.h"
18#include "xfs_attr.h"
19#include "xfs_trans_space.h"
20#include "xfs_trans.h"
21#include "xfs_buf_item.h"
22#include "xfs_inode_item.h"
23#include "xfs_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}
1/*
2 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
3 * All Rights Reserved.
4 *
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
8 *
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
13 *
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
17 */
18#include <linux/log2.h>
19
20#include "xfs.h"
21#include "xfs_fs.h"
22#include "xfs_shared.h"
23#include "xfs_format.h"
24#include "xfs_log_format.h"
25#include "xfs_trans_resv.h"
26#include "xfs_sb.h"
27#include "xfs_mount.h"
28#include "xfs_inode.h"
29#include "xfs_da_format.h"
30#include "xfs_da_btree.h"
31#include "xfs_dir2.h"
32#include "xfs_attr_sf.h"
33#include "xfs_attr.h"
34#include "xfs_trans_space.h"
35#include "xfs_trans.h"
36#include "xfs_buf_item.h"
37#include "xfs_inode_item.h"
38#include "xfs_ialloc.h"
39#include "xfs_bmap.h"
40#include "xfs_bmap_util.h"
41#include "xfs_error.h"
42#include "xfs_quota.h"
43#include "xfs_filestream.h"
44#include "xfs_cksum.h"
45#include "xfs_trace.h"
46#include "xfs_icache.h"
47#include "xfs_symlink.h"
48#include "xfs_trans_priv.h"
49#include "xfs_log.h"
50#include "xfs_bmap_btree.h"
51
52kmem_zone_t *xfs_inode_zone;
53
54/*
55 * Used in xfs_itruncate_extents(). This is the maximum number of extents
56 * freed from a file in a single transaction.
57 */
58#define XFS_ITRUNC_MAX_EXTENTS 2
59
60STATIC int xfs_iflush_int(struct xfs_inode *, struct xfs_buf *);
61STATIC int xfs_iunlink(struct xfs_trans *, struct xfs_inode *);
62STATIC int xfs_iunlink_remove(struct xfs_trans *, struct xfs_inode *);
63
64/*
65 * helper function to extract extent size hint from inode
66 */
67xfs_extlen_t
68xfs_get_extsz_hint(
69 struct xfs_inode *ip)
70{
71 if ((ip->i_d.di_flags & XFS_DIFLAG_EXTSIZE) && ip->i_d.di_extsize)
72 return ip->i_d.di_extsize;
73 if (XFS_IS_REALTIME_INODE(ip))
74 return ip->i_mount->m_sb.sb_rextsize;
75 return 0;
76}
77
78/*
79 * These two are wrapper routines around the xfs_ilock() routine used to
80 * centralize some grungy code. They are used in places that wish to lock the
81 * inode solely for reading the extents. The reason these places can't just
82 * call xfs_ilock(ip, XFS_ILOCK_SHARED) is that the inode lock also guards to
83 * bringing in of the extents from disk for a file in b-tree format. If the
84 * inode is in b-tree format, then we need to lock the inode exclusively until
85 * the extents are read in. Locking it exclusively all the time would limit
86 * our parallelism unnecessarily, though. What we do instead is check to see
87 * if the extents have been read in yet, and only lock the inode exclusively
88 * if they have not.
89 *
90 * The functions return a value which should be given to the corresponding
91 * xfs_iunlock() call.
92 */
93uint
94xfs_ilock_data_map_shared(
95 struct xfs_inode *ip)
96{
97 uint lock_mode = XFS_ILOCK_SHARED;
98
99 if (ip->i_d.di_format == XFS_DINODE_FMT_BTREE &&
100 (ip->i_df.if_flags & XFS_IFEXTENTS) == 0)
101 lock_mode = XFS_ILOCK_EXCL;
102 xfs_ilock(ip, lock_mode);
103 return lock_mode;
104}
105
106uint
107xfs_ilock_attr_map_shared(
108 struct xfs_inode *ip)
109{
110 uint lock_mode = XFS_ILOCK_SHARED;
111
112 if (ip->i_d.di_aformat == XFS_DINODE_FMT_BTREE &&
113 (ip->i_afp->if_flags & XFS_IFEXTENTS) == 0)
114 lock_mode = XFS_ILOCK_EXCL;
115 xfs_ilock(ip, lock_mode);
116 return lock_mode;
117}
118
119/*
120 * The xfs inode contains 3 multi-reader locks: the i_iolock the i_mmap_lock and
121 * the i_lock. This routine allows various combinations of the locks to be
122 * obtained.
123 *
124 * The 3 locks should always be ordered so that the IO lock is obtained first,
125 * the mmap lock second and the ilock last in order to prevent deadlock.
126 *
127 * Basic locking order:
128 *
129 * i_iolock -> i_mmap_lock -> page_lock -> i_ilock
130 *
131 * mmap_sem locking order:
132 *
133 * i_iolock -> page lock -> mmap_sem
134 * mmap_sem -> i_mmap_lock -> page_lock
135 *
136 * The difference in mmap_sem locking order mean that we cannot hold the
137 * i_mmap_lock over syscall based read(2)/write(2) based IO. These IO paths can
138 * fault in pages during copy in/out (for buffered IO) or require the mmap_sem
139 * in get_user_pages() to map the user pages into the kernel address space for
140 * direct IO. Similarly the i_iolock cannot be taken inside a page fault because
141 * page faults already hold the mmap_sem.
142 *
143 * Hence to serialise fully against both syscall and mmap based IO, we need to
144 * take both the i_iolock and the i_mmap_lock. These locks should *only* be both
145 * taken in places where we need to invalidate the page cache in a race
146 * free manner (e.g. truncate, hole punch and other extent manipulation
147 * functions).
148 */
149void
150xfs_ilock(
151 xfs_inode_t *ip,
152 uint lock_flags)
153{
154 trace_xfs_ilock(ip, lock_flags, _RET_IP_);
155
156 /*
157 * You can't set both SHARED and EXCL for the same lock,
158 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
159 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
160 */
161 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
162 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
163 ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
164 (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
165 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
166 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
167 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
168
169 if (lock_flags & XFS_IOLOCK_EXCL)
170 mrupdate_nested(&ip->i_iolock, XFS_IOLOCK_DEP(lock_flags));
171 else if (lock_flags & XFS_IOLOCK_SHARED)
172 mraccess_nested(&ip->i_iolock, XFS_IOLOCK_DEP(lock_flags));
173
174 if (lock_flags & XFS_MMAPLOCK_EXCL)
175 mrupdate_nested(&ip->i_mmaplock, XFS_MMAPLOCK_DEP(lock_flags));
176 else if (lock_flags & XFS_MMAPLOCK_SHARED)
177 mraccess_nested(&ip->i_mmaplock, XFS_MMAPLOCK_DEP(lock_flags));
178
179 if (lock_flags & XFS_ILOCK_EXCL)
180 mrupdate_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
181 else if (lock_flags & XFS_ILOCK_SHARED)
182 mraccess_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
183}
184
185/*
186 * This is just like xfs_ilock(), except that the caller
187 * is guaranteed not to sleep. It returns 1 if it gets
188 * the requested locks and 0 otherwise. If the IO lock is
189 * obtained but the inode lock cannot be, then the IO lock
190 * is dropped before returning.
191 *
192 * ip -- the inode being locked
193 * lock_flags -- this parameter indicates the inode's locks to be
194 * to be locked. See the comment for xfs_ilock() for a list
195 * of valid values.
196 */
197int
198xfs_ilock_nowait(
199 xfs_inode_t *ip,
200 uint lock_flags)
201{
202 trace_xfs_ilock_nowait(ip, lock_flags, _RET_IP_);
203
204 /*
205 * You can't set both SHARED and EXCL for the same lock,
206 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
207 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
208 */
209 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
210 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
211 ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
212 (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
213 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
214 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
215 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
216
217 if (lock_flags & XFS_IOLOCK_EXCL) {
218 if (!mrtryupdate(&ip->i_iolock))
219 goto out;
220 } else if (lock_flags & XFS_IOLOCK_SHARED) {
221 if (!mrtryaccess(&ip->i_iolock))
222 goto out;
223 }
224
225 if (lock_flags & XFS_MMAPLOCK_EXCL) {
226 if (!mrtryupdate(&ip->i_mmaplock))
227 goto out_undo_iolock;
228 } else if (lock_flags & XFS_MMAPLOCK_SHARED) {
229 if (!mrtryaccess(&ip->i_mmaplock))
230 goto out_undo_iolock;
231 }
232
233 if (lock_flags & XFS_ILOCK_EXCL) {
234 if (!mrtryupdate(&ip->i_lock))
235 goto out_undo_mmaplock;
236 } else if (lock_flags & XFS_ILOCK_SHARED) {
237 if (!mrtryaccess(&ip->i_lock))
238 goto out_undo_mmaplock;
239 }
240 return 1;
241
242out_undo_mmaplock:
243 if (lock_flags & XFS_MMAPLOCK_EXCL)
244 mrunlock_excl(&ip->i_mmaplock);
245 else if (lock_flags & XFS_MMAPLOCK_SHARED)
246 mrunlock_shared(&ip->i_mmaplock);
247out_undo_iolock:
248 if (lock_flags & XFS_IOLOCK_EXCL)
249 mrunlock_excl(&ip->i_iolock);
250 else if (lock_flags & XFS_IOLOCK_SHARED)
251 mrunlock_shared(&ip->i_iolock);
252out:
253 return 0;
254}
255
256/*
257 * xfs_iunlock() is used to drop the inode locks acquired with
258 * xfs_ilock() and xfs_ilock_nowait(). The caller must pass
259 * in the flags given to xfs_ilock() or xfs_ilock_nowait() so
260 * that we know which locks to drop.
261 *
262 * ip -- the inode being unlocked
263 * lock_flags -- this parameter indicates the inode's locks to be
264 * to be unlocked. See the comment for xfs_ilock() for a list
265 * of valid values for this parameter.
266 *
267 */
268void
269xfs_iunlock(
270 xfs_inode_t *ip,
271 uint lock_flags)
272{
273 /*
274 * You can't set both SHARED and EXCL for the same lock,
275 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
276 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
277 */
278 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
279 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
280 ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
281 (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
282 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
283 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
284 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
285 ASSERT(lock_flags != 0);
286
287 if (lock_flags & XFS_IOLOCK_EXCL)
288 mrunlock_excl(&ip->i_iolock);
289 else if (lock_flags & XFS_IOLOCK_SHARED)
290 mrunlock_shared(&ip->i_iolock);
291
292 if (lock_flags & XFS_MMAPLOCK_EXCL)
293 mrunlock_excl(&ip->i_mmaplock);
294 else if (lock_flags & XFS_MMAPLOCK_SHARED)
295 mrunlock_shared(&ip->i_mmaplock);
296
297 if (lock_flags & XFS_ILOCK_EXCL)
298 mrunlock_excl(&ip->i_lock);
299 else if (lock_flags & XFS_ILOCK_SHARED)
300 mrunlock_shared(&ip->i_lock);
301
302 trace_xfs_iunlock(ip, lock_flags, _RET_IP_);
303}
304
305/*
306 * give up write locks. the i/o lock cannot be held nested
307 * if it is being demoted.
308 */
309void
310xfs_ilock_demote(
311 xfs_inode_t *ip,
312 uint lock_flags)
313{
314 ASSERT(lock_flags & (XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL));
315 ASSERT((lock_flags &
316 ~(XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL)) == 0);
317
318 if (lock_flags & XFS_ILOCK_EXCL)
319 mrdemote(&ip->i_lock);
320 if (lock_flags & XFS_MMAPLOCK_EXCL)
321 mrdemote(&ip->i_mmaplock);
322 if (lock_flags & XFS_IOLOCK_EXCL)
323 mrdemote(&ip->i_iolock);
324
325 trace_xfs_ilock_demote(ip, lock_flags, _RET_IP_);
326}
327
328#if defined(DEBUG) || defined(XFS_WARN)
329int
330xfs_isilocked(
331 xfs_inode_t *ip,
332 uint lock_flags)
333{
334 if (lock_flags & (XFS_ILOCK_EXCL|XFS_ILOCK_SHARED)) {
335 if (!(lock_flags & XFS_ILOCK_SHARED))
336 return !!ip->i_lock.mr_writer;
337 return rwsem_is_locked(&ip->i_lock.mr_lock);
338 }
339
340 if (lock_flags & (XFS_MMAPLOCK_EXCL|XFS_MMAPLOCK_SHARED)) {
341 if (!(lock_flags & XFS_MMAPLOCK_SHARED))
342 return !!ip->i_mmaplock.mr_writer;
343 return rwsem_is_locked(&ip->i_mmaplock.mr_lock);
344 }
345
346 if (lock_flags & (XFS_IOLOCK_EXCL|XFS_IOLOCK_SHARED)) {
347 if (!(lock_flags & XFS_IOLOCK_SHARED))
348 return !!ip->i_iolock.mr_writer;
349 return rwsem_is_locked(&ip->i_iolock.mr_lock);
350 }
351
352 ASSERT(0);
353 return 0;
354}
355#endif
356
357#ifdef DEBUG
358int xfs_locked_n;
359int xfs_small_retries;
360int xfs_middle_retries;
361int xfs_lots_retries;
362int xfs_lock_delays;
363#endif
364
365/*
366 * xfs_lockdep_subclass_ok() is only used in an ASSERT, so is only called when
367 * DEBUG or XFS_WARN is set. And MAX_LOCKDEP_SUBCLASSES is then only defined
368 * when CONFIG_LOCKDEP is set. Hence the complex define below to avoid build
369 * errors and warnings.
370 */
371#if (defined(DEBUG) || defined(XFS_WARN)) && defined(CONFIG_LOCKDEP)
372static bool
373xfs_lockdep_subclass_ok(
374 int subclass)
375{
376 return subclass < MAX_LOCKDEP_SUBCLASSES;
377}
378#else
379#define xfs_lockdep_subclass_ok(subclass) (true)
380#endif
381
382/*
383 * Bump the subclass so xfs_lock_inodes() acquires each lock with a different
384 * value. This can be called for any type of inode lock combination, including
385 * parent locking. Care must be taken to ensure we don't overrun the subclass
386 * storage fields in the class mask we build.
387 */
388static inline int
389xfs_lock_inumorder(int lock_mode, int subclass)
390{
391 int class = 0;
392
393 ASSERT(!(lock_mode & (XFS_ILOCK_PARENT | XFS_ILOCK_RTBITMAP |
394 XFS_ILOCK_RTSUM)));
395 ASSERT(xfs_lockdep_subclass_ok(subclass));
396
397 if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)) {
398 ASSERT(subclass <= XFS_IOLOCK_MAX_SUBCLASS);
399 ASSERT(xfs_lockdep_subclass_ok(subclass +
400 XFS_IOLOCK_PARENT_VAL));
401 class += subclass << XFS_IOLOCK_SHIFT;
402 if (lock_mode & XFS_IOLOCK_PARENT)
403 class += XFS_IOLOCK_PARENT_VAL << XFS_IOLOCK_SHIFT;
404 }
405
406 if (lock_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) {
407 ASSERT(subclass <= XFS_MMAPLOCK_MAX_SUBCLASS);
408 class += subclass << XFS_MMAPLOCK_SHIFT;
409 }
410
411 if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) {
412 ASSERT(subclass <= XFS_ILOCK_MAX_SUBCLASS);
413 class += subclass << XFS_ILOCK_SHIFT;
414 }
415
416 return (lock_mode & ~XFS_LOCK_SUBCLASS_MASK) | class;
417}
418
419/*
420 * The following routine will lock n inodes in exclusive mode. We assume the
421 * caller calls us with the inodes in i_ino order.
422 *
423 * We need to detect deadlock where an inode that we lock is in the AIL and we
424 * start waiting for another inode that is locked by a thread in a long running
425 * transaction (such as truncate). This can result in deadlock since the long
426 * running trans might need to wait for the inode we just locked in order to
427 * push the tail and free space in the log.
428 *
429 * xfs_lock_inodes() can only be used to lock one type of lock at a time -
430 * the iolock, the mmaplock or the ilock, but not more than one at a time. If we
431 * lock more than one at a time, lockdep will report false positives saying we
432 * have violated locking orders.
433 */
434void
435xfs_lock_inodes(
436 xfs_inode_t **ips,
437 int inodes,
438 uint lock_mode)
439{
440 int attempts = 0, i, j, try_lock;
441 xfs_log_item_t *lp;
442
443 /*
444 * Currently supports between 2 and 5 inodes with exclusive locking. We
445 * support an arbitrary depth of locking here, but absolute limits on
446 * inodes depend on the the type of locking and the limits placed by
447 * lockdep annotations in xfs_lock_inumorder. These are all checked by
448 * the asserts.
449 */
450 ASSERT(ips && inodes >= 2 && inodes <= 5);
451 ASSERT(lock_mode & (XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL |
452 XFS_ILOCK_EXCL));
453 ASSERT(!(lock_mode & (XFS_IOLOCK_SHARED | XFS_MMAPLOCK_SHARED |
454 XFS_ILOCK_SHARED)));
455 ASSERT(!(lock_mode & XFS_IOLOCK_EXCL) ||
456 inodes <= XFS_IOLOCK_MAX_SUBCLASS + 1);
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 = (xfs_log_item_t *)ips[j]->i_itemp;
483 if (lp && (lp->li_flags & XFS_LI_IN_AIL))
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#ifdef DEBUG
524 xfs_lock_delays++;
525#endif
526 }
527 i = 0;
528 try_lock = 0;
529 goto again;
530 }
531
532#ifdef DEBUG
533 if (attempts) {
534 if (attempts < 5) xfs_small_retries++;
535 else if (attempts < 100) xfs_middle_retries++;
536 else xfs_lots_retries++;
537 } else {
538 xfs_locked_n++;
539 }
540#endif
541}
542
543/*
544 * xfs_lock_two_inodes() can only be used to lock one type of lock at a time -
545 * the iolock, the mmaplock or the ilock, but not more than one at a time. If we
546 * lock more than one at a time, lockdep will report false positives saying we
547 * have violated locking orders.
548 */
549void
550xfs_lock_two_inodes(
551 xfs_inode_t *ip0,
552 xfs_inode_t *ip1,
553 uint lock_mode)
554{
555 xfs_inode_t *temp;
556 int attempts = 0;
557 xfs_log_item_t *lp;
558
559 if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)) {
560 ASSERT(!(lock_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)));
561 ASSERT(!(lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
562 } else if (lock_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL))
563 ASSERT(!(lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
564
565 ASSERT(ip0->i_ino != ip1->i_ino);
566
567 if (ip0->i_ino > ip1->i_ino) {
568 temp = ip0;
569 ip0 = ip1;
570 ip1 = temp;
571 }
572
573 again:
574 xfs_ilock(ip0, xfs_lock_inumorder(lock_mode, 0));
575
576 /*
577 * If the first lock we have locked is in the AIL, we must TRY to get
578 * the second lock. If we can't get it, we must release the first one
579 * and try again.
580 */
581 lp = (xfs_log_item_t *)ip0->i_itemp;
582 if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
583 if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(lock_mode, 1))) {
584 xfs_iunlock(ip0, lock_mode);
585 if ((++attempts % 5) == 0)
586 delay(1); /* Don't just spin the CPU */
587 goto again;
588 }
589 } else {
590 xfs_ilock(ip1, xfs_lock_inumorder(lock_mode, 1));
591 }
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.wait, TASK_UNINTERRUPTIBLE);
604 if (xfs_isiflocked(ip))
605 io_schedule();
606 } while (!xfs_iflock_nowait(ip));
607
608 finish_wait(wq, &wait.wait);
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 }
654
655 if (has_attr)
656 flags |= FS_XFLAG_HASATTR;
657
658 return flags;
659}
660
661uint
662xfs_ip2xflags(
663 struct xfs_inode *ip)
664{
665 struct xfs_icdinode *dic = &ip->i_d;
666
667 return _xfs_dic2xflags(dic->di_flags, dic->di_flags2, XFS_IFORK_Q(ip));
668}
669
670uint
671xfs_dic2xflags(
672 struct xfs_dinode *dip)
673{
674 return _xfs_dic2xflags(be16_to_cpu(dip->di_flags),
675 be64_to_cpu(dip->di_flags2), XFS_DFORK_Q(dip));
676}
677
678/*
679 * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
680 * is allowed, otherwise it has to be an exact match. If a CI match is found,
681 * ci_name->name will point to a the actual name (caller must free) or
682 * will be set to NULL if an exact match is found.
683 */
684int
685xfs_lookup(
686 xfs_inode_t *dp,
687 struct xfs_name *name,
688 xfs_inode_t **ipp,
689 struct xfs_name *ci_name)
690{
691 xfs_ino_t inum;
692 int error;
693
694 trace_xfs_lookup(dp, name);
695
696 if (XFS_FORCED_SHUTDOWN(dp->i_mount))
697 return -EIO;
698
699 xfs_ilock(dp, XFS_IOLOCK_SHARED);
700 error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
701 if (error)
702 goto out_unlock;
703
704 error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp);
705 if (error)
706 goto out_free_name;
707
708 xfs_iunlock(dp, XFS_IOLOCK_SHARED);
709 return 0;
710
711out_free_name:
712 if (ci_name)
713 kmem_free(ci_name->name);
714out_unlock:
715 xfs_iunlock(dp, XFS_IOLOCK_SHARED);
716 *ipp = NULL;
717 return error;
718}
719
720/*
721 * Allocate an inode on disk and return a copy of its in-core version.
722 * The in-core inode is locked exclusively. Set mode, nlink, and rdev
723 * appropriately within the inode. The uid and gid for the inode are
724 * set according to the contents of the given cred structure.
725 *
726 * Use xfs_dialloc() to allocate the on-disk inode. If xfs_dialloc()
727 * has a free inode available, call xfs_iget() to obtain the in-core
728 * version of the allocated inode. Finally, fill in the inode and
729 * log its initial contents. In this case, ialloc_context would be
730 * set to NULL.
731 *
732 * If xfs_dialloc() does not have an available inode, it will replenish
733 * its supply by doing an allocation. Since we can only do one
734 * allocation within a transaction without deadlocks, we must commit
735 * the current transaction before returning the inode itself.
736 * In this case, therefore, we will set ialloc_context and return.
737 * The caller should then commit the current transaction, start a new
738 * transaction, and call xfs_ialloc() again to actually get the inode.
739 *
740 * To ensure that some other process does not grab the inode that
741 * was allocated during the first call to xfs_ialloc(), this routine
742 * also returns the [locked] bp pointing to the head of the freelist
743 * as ialloc_context. The caller should hold this buffer across
744 * the commit and pass it back into this routine on the second call.
745 *
746 * If we are allocating quota inodes, we do not have a parent inode
747 * to attach to or associate with (i.e. pip == NULL) because they
748 * are not linked into the directory structure - they are attached
749 * directly to the superblock - and so have no parent.
750 */
751int
752xfs_ialloc(
753 xfs_trans_t *tp,
754 xfs_inode_t *pip,
755 umode_t mode,
756 xfs_nlink_t nlink,
757 xfs_dev_t rdev,
758 prid_t prid,
759 int okalloc,
760 xfs_buf_t **ialloc_context,
761 xfs_inode_t **ipp)
762{
763 struct xfs_mount *mp = tp->t_mountp;
764 xfs_ino_t ino;
765 xfs_inode_t *ip;
766 uint flags;
767 int error;
768 struct timespec tv;
769 struct inode *inode;
770
771 /*
772 * Call the space management code to pick
773 * the on-disk inode to be allocated.
774 */
775 error = xfs_dialloc(tp, pip ? pip->i_ino : 0, mode, okalloc,
776 ialloc_context, &ino);
777 if (error)
778 return error;
779 if (*ialloc_context || ino == NULLFSINO) {
780 *ipp = NULL;
781 return 0;
782 }
783 ASSERT(*ialloc_context == NULL);
784
785 /*
786 * Get the in-core inode with the lock held exclusively.
787 * This is because we're setting fields here we need
788 * to prevent others from looking at until we're done.
789 */
790 error = xfs_iget(mp, tp, ino, XFS_IGET_CREATE,
791 XFS_ILOCK_EXCL, &ip);
792 if (error)
793 return error;
794 ASSERT(ip != NULL);
795 inode = VFS_I(ip);
796
797 /*
798 * We always convert v1 inodes to v2 now - we only support filesystems
799 * with >= v2 inode capability, so there is no reason for ever leaving
800 * an inode in v1 format.
801 */
802 if (ip->i_d.di_version == 1)
803 ip->i_d.di_version = 2;
804
805 inode->i_mode = mode;
806 set_nlink(inode, nlink);
807 ip->i_d.di_uid = xfs_kuid_to_uid(current_fsuid());
808 ip->i_d.di_gid = xfs_kgid_to_gid(current_fsgid());
809 xfs_set_projid(ip, prid);
810
811 if (pip && XFS_INHERIT_GID(pip)) {
812 ip->i_d.di_gid = pip->i_d.di_gid;
813 if ((VFS_I(pip)->i_mode & S_ISGID) && S_ISDIR(mode))
814 inode->i_mode |= S_ISGID;
815 }
816
817 /*
818 * If the group ID of the new file does not match the effective group
819 * ID or one of the supplementary group IDs, the S_ISGID bit is cleared
820 * (and only if the irix_sgid_inherit compatibility variable is set).
821 */
822 if ((irix_sgid_inherit) &&
823 (inode->i_mode & S_ISGID) &&
824 (!in_group_p(xfs_gid_to_kgid(ip->i_d.di_gid))))
825 inode->i_mode &= ~S_ISGID;
826
827 ip->i_d.di_size = 0;
828 ip->i_d.di_nextents = 0;
829 ASSERT(ip->i_d.di_nblocks == 0);
830
831 tv = current_fs_time(mp->m_super);
832 inode->i_mtime = tv;
833 inode->i_atime = tv;
834 inode->i_ctime = tv;
835
836 ip->i_d.di_extsize = 0;
837 ip->i_d.di_dmevmask = 0;
838 ip->i_d.di_dmstate = 0;
839 ip->i_d.di_flags = 0;
840
841 if (ip->i_d.di_version == 3) {
842 inode->i_version = 1;
843 ip->i_d.di_flags2 = 0;
844 ip->i_d.di_crtime.t_sec = (__int32_t)tv.tv_sec;
845 ip->i_d.di_crtime.t_nsec = (__int32_t)tv.tv_nsec;
846 }
847
848
849 flags = XFS_ILOG_CORE;
850 switch (mode & S_IFMT) {
851 case S_IFIFO:
852 case S_IFCHR:
853 case S_IFBLK:
854 case S_IFSOCK:
855 ip->i_d.di_format = XFS_DINODE_FMT_DEV;
856 ip->i_df.if_u2.if_rdev = rdev;
857 ip->i_df.if_flags = 0;
858 flags |= XFS_ILOG_DEV;
859 break;
860 case S_IFREG:
861 case S_IFDIR:
862 if (pip && (pip->i_d.di_flags & XFS_DIFLAG_ANY)) {
863 uint64_t di_flags2 = 0;
864 uint di_flags = 0;
865
866 if (S_ISDIR(mode)) {
867 if (pip->i_d.di_flags & XFS_DIFLAG_RTINHERIT)
868 di_flags |= XFS_DIFLAG_RTINHERIT;
869 if (pip->i_d.di_flags & XFS_DIFLAG_EXTSZINHERIT) {
870 di_flags |= XFS_DIFLAG_EXTSZINHERIT;
871 ip->i_d.di_extsize = pip->i_d.di_extsize;
872 }
873 if (pip->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
874 di_flags |= XFS_DIFLAG_PROJINHERIT;
875 } else if (S_ISREG(mode)) {
876 if (pip->i_d.di_flags & XFS_DIFLAG_RTINHERIT)
877 di_flags |= XFS_DIFLAG_REALTIME;
878 if (pip->i_d.di_flags & XFS_DIFLAG_EXTSZINHERIT) {
879 di_flags |= XFS_DIFLAG_EXTSIZE;
880 ip->i_d.di_extsize = pip->i_d.di_extsize;
881 }
882 }
883 if ((pip->i_d.di_flags & XFS_DIFLAG_NOATIME) &&
884 xfs_inherit_noatime)
885 di_flags |= XFS_DIFLAG_NOATIME;
886 if ((pip->i_d.di_flags & XFS_DIFLAG_NODUMP) &&
887 xfs_inherit_nodump)
888 di_flags |= XFS_DIFLAG_NODUMP;
889 if ((pip->i_d.di_flags & XFS_DIFLAG_SYNC) &&
890 xfs_inherit_sync)
891 di_flags |= XFS_DIFLAG_SYNC;
892 if ((pip->i_d.di_flags & XFS_DIFLAG_NOSYMLINKS) &&
893 xfs_inherit_nosymlinks)
894 di_flags |= XFS_DIFLAG_NOSYMLINKS;
895 if ((pip->i_d.di_flags & XFS_DIFLAG_NODEFRAG) &&
896 xfs_inherit_nodefrag)
897 di_flags |= XFS_DIFLAG_NODEFRAG;
898 if (pip->i_d.di_flags & XFS_DIFLAG_FILESTREAM)
899 di_flags |= XFS_DIFLAG_FILESTREAM;
900 if (pip->i_d.di_flags2 & XFS_DIFLAG2_DAX)
901 di_flags2 |= XFS_DIFLAG2_DAX;
902
903 ip->i_d.di_flags |= di_flags;
904 ip->i_d.di_flags2 |= di_flags2;
905 }
906 /* FALLTHROUGH */
907 case S_IFLNK:
908 ip->i_d.di_format = XFS_DINODE_FMT_EXTENTS;
909 ip->i_df.if_flags = XFS_IFEXTENTS;
910 ip->i_df.if_bytes = ip->i_df.if_real_bytes = 0;
911 ip->i_df.if_u1.if_extents = NULL;
912 break;
913 default:
914 ASSERT(0);
915 }
916 /*
917 * Attribute fork settings for new inode.
918 */
919 ip->i_d.di_aformat = XFS_DINODE_FMT_EXTENTS;
920 ip->i_d.di_anextents = 0;
921
922 /*
923 * Log the new values stuffed into the inode.
924 */
925 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
926 xfs_trans_log_inode(tp, ip, flags);
927
928 /* now that we have an i_mode we can setup the inode structure */
929 xfs_setup_inode(ip);
930
931 *ipp = ip;
932 return 0;
933}
934
935/*
936 * Allocates a new inode from disk and return a pointer to the
937 * incore copy. This routine will internally commit the current
938 * transaction and allocate a new one if the Space Manager needed
939 * to do an allocation to replenish the inode free-list.
940 *
941 * This routine is designed to be called from xfs_create and
942 * xfs_create_dir.
943 *
944 */
945int
946xfs_dir_ialloc(
947 xfs_trans_t **tpp, /* input: current transaction;
948 output: may be a new transaction. */
949 xfs_inode_t *dp, /* directory within whose allocate
950 the inode. */
951 umode_t mode,
952 xfs_nlink_t nlink,
953 xfs_dev_t rdev,
954 prid_t prid, /* project id */
955 int okalloc, /* ok to allocate new space */
956 xfs_inode_t **ipp, /* pointer to inode; it will be
957 locked. */
958 int *committed)
959
960{
961 xfs_trans_t *tp;
962 xfs_inode_t *ip;
963 xfs_buf_t *ialloc_context = NULL;
964 int code;
965 void *dqinfo;
966 uint tflags;
967
968 tp = *tpp;
969 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
970
971 /*
972 * xfs_ialloc will return a pointer to an incore inode if
973 * the Space Manager has an available inode on the free
974 * list. Otherwise, it will do an allocation and replenish
975 * the freelist. Since we can only do one allocation per
976 * transaction without deadlocks, we will need to commit the
977 * current transaction and start a new one. We will then
978 * need to call xfs_ialloc again to get the inode.
979 *
980 * If xfs_ialloc did an allocation to replenish the freelist,
981 * it returns the bp containing the head of the freelist as
982 * ialloc_context. We will hold a lock on it across the
983 * transaction commit so that no other process can steal
984 * the inode(s) that we've just allocated.
985 */
986 code = xfs_ialloc(tp, dp, mode, nlink, rdev, prid, okalloc,
987 &ialloc_context, &ip);
988
989 /*
990 * Return an error if we were unable to allocate a new inode.
991 * This should only happen if we run out of space on disk or
992 * encounter a disk error.
993 */
994 if (code) {
995 *ipp = NULL;
996 return code;
997 }
998 if (!ialloc_context && !ip) {
999 *ipp = NULL;
1000 return -ENOSPC;
1001 }
1002
1003 /*
1004 * If the AGI buffer is non-NULL, then we were unable to get an
1005 * inode in one operation. We need to commit the current
1006 * transaction and call xfs_ialloc() again. It is guaranteed
1007 * to succeed the second time.
1008 */
1009 if (ialloc_context) {
1010 /*
1011 * Normally, xfs_trans_commit releases all the locks.
1012 * We call bhold to hang on to the ialloc_context across
1013 * the commit. Holding this buffer prevents any other
1014 * processes from doing any allocations in this
1015 * allocation group.
1016 */
1017 xfs_trans_bhold(tp, ialloc_context);
1018
1019 /*
1020 * We want the quota changes to be associated with the next
1021 * transaction, NOT this one. So, detach the dqinfo from this
1022 * and attach it to the next transaction.
1023 */
1024 dqinfo = NULL;
1025 tflags = 0;
1026 if (tp->t_dqinfo) {
1027 dqinfo = (void *)tp->t_dqinfo;
1028 tp->t_dqinfo = NULL;
1029 tflags = tp->t_flags & XFS_TRANS_DQ_DIRTY;
1030 tp->t_flags &= ~(XFS_TRANS_DQ_DIRTY);
1031 }
1032
1033 code = xfs_trans_roll(&tp, 0);
1034 if (committed != NULL)
1035 *committed = 1;
1036
1037 /*
1038 * Re-attach the quota info that we detached from prev trx.
1039 */
1040 if (dqinfo) {
1041 tp->t_dqinfo = dqinfo;
1042 tp->t_flags |= tflags;
1043 }
1044
1045 if (code) {
1046 xfs_buf_relse(ialloc_context);
1047 *tpp = tp;
1048 *ipp = NULL;
1049 return code;
1050 }
1051 xfs_trans_bjoin(tp, ialloc_context);
1052
1053 /*
1054 * Call ialloc again. Since we've locked out all
1055 * other allocations in this allocation group,
1056 * this call should always succeed.
1057 */
1058 code = xfs_ialloc(tp, dp, mode, nlink, rdev, prid,
1059 okalloc, &ialloc_context, &ip);
1060
1061 /*
1062 * If we get an error at this point, return to the caller
1063 * so that the current transaction can be aborted.
1064 */
1065 if (code) {
1066 *tpp = tp;
1067 *ipp = NULL;
1068 return code;
1069 }
1070 ASSERT(!ialloc_context && ip);
1071
1072 } else {
1073 if (committed != NULL)
1074 *committed = 0;
1075 }
1076
1077 *ipp = ip;
1078 *tpp = tp;
1079
1080 return 0;
1081}
1082
1083/*
1084 * Decrement the link count on an inode & log the change. If this causes the
1085 * link count to go to zero, move the inode to AGI unlinked list so that it can
1086 * be freed when the last active reference goes away via xfs_inactive().
1087 */
1088int /* error */
1089xfs_droplink(
1090 xfs_trans_t *tp,
1091 xfs_inode_t *ip)
1092{
1093 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
1094
1095 drop_nlink(VFS_I(ip));
1096 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1097
1098 if (VFS_I(ip)->i_nlink)
1099 return 0;
1100
1101 return xfs_iunlink(tp, ip);
1102}
1103
1104/*
1105 * Increment the link count on an inode & log the change.
1106 */
1107int
1108xfs_bumplink(
1109 xfs_trans_t *tp,
1110 xfs_inode_t *ip)
1111{
1112 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
1113
1114 ASSERT(ip->i_d.di_version > 1);
1115 inc_nlink(VFS_I(ip));
1116 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1117 return 0;
1118}
1119
1120int
1121xfs_create(
1122 xfs_inode_t *dp,
1123 struct xfs_name *name,
1124 umode_t mode,
1125 xfs_dev_t rdev,
1126 xfs_inode_t **ipp)
1127{
1128 int is_dir = S_ISDIR(mode);
1129 struct xfs_mount *mp = dp->i_mount;
1130 struct xfs_inode *ip = NULL;
1131 struct xfs_trans *tp = NULL;
1132 int error;
1133 xfs_bmap_free_t free_list;
1134 xfs_fsblock_t first_block;
1135 bool unlock_dp_on_error = false;
1136 prid_t prid;
1137 struct xfs_dquot *udqp = NULL;
1138 struct xfs_dquot *gdqp = NULL;
1139 struct xfs_dquot *pdqp = NULL;
1140 struct xfs_trans_res *tres;
1141 uint resblks;
1142
1143 trace_xfs_create(dp, name);
1144
1145 if (XFS_FORCED_SHUTDOWN(mp))
1146 return -EIO;
1147
1148 prid = xfs_get_initial_prid(dp);
1149
1150 /*
1151 * Make sure that we have allocated dquot(s) on disk.
1152 */
1153 error = xfs_qm_vop_dqalloc(dp, xfs_kuid_to_uid(current_fsuid()),
1154 xfs_kgid_to_gid(current_fsgid()), prid,
1155 XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
1156 &udqp, &gdqp, &pdqp);
1157 if (error)
1158 return error;
1159
1160 if (is_dir) {
1161 rdev = 0;
1162 resblks = XFS_MKDIR_SPACE_RES(mp, name->len);
1163 tres = &M_RES(mp)->tr_mkdir;
1164 tp = xfs_trans_alloc(mp, XFS_TRANS_MKDIR);
1165 } else {
1166 resblks = XFS_CREATE_SPACE_RES(mp, name->len);
1167 tres = &M_RES(mp)->tr_create;
1168 tp = xfs_trans_alloc(mp, XFS_TRANS_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_reserve(tp, tres, resblks, 0);
1178 if (error == -ENOSPC) {
1179 /* flush outstanding delalloc blocks and retry */
1180 xfs_flush_inodes(mp);
1181 error = xfs_trans_reserve(tp, tres, resblks, 0);
1182 }
1183 if (error == -ENOSPC) {
1184 /* No space at all so try a "no-allocation" reservation */
1185 resblks = 0;
1186 error = xfs_trans_reserve(tp, tres, 0, 0);
1187 }
1188 if (error)
1189 goto out_trans_cancel;
1190
1191
1192 xfs_ilock(dp, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL |
1193 XFS_IOLOCK_PARENT | XFS_ILOCK_PARENT);
1194 unlock_dp_on_error = true;
1195
1196 xfs_bmap_init(&free_list, &first_block);
1197
1198 /*
1199 * Reserve disk quota and the inode.
1200 */
1201 error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp,
1202 pdqp, resblks, 1, 0);
1203 if (error)
1204 goto out_trans_cancel;
1205
1206 if (!resblks) {
1207 error = xfs_dir_canenter(tp, dp, name);
1208 if (error)
1209 goto out_trans_cancel;
1210 }
1211
1212 /*
1213 * A newly created regular or special file just has one directory
1214 * entry pointing to them, but a directory also the "." entry
1215 * pointing to itself.
1216 */
1217 error = xfs_dir_ialloc(&tp, dp, mode, is_dir ? 2 : 1, rdev,
1218 prid, resblks > 0, &ip, NULL);
1219 if (error)
1220 goto out_trans_cancel;
1221
1222 /*
1223 * Now we join the directory inode to the transaction. We do not do it
1224 * earlier because xfs_dir_ialloc might commit the previous transaction
1225 * (and release all the locks). An error from here on will result in
1226 * the transaction cancel unlocking dp so don't do it explicitly in the
1227 * error path.
1228 */
1229 xfs_trans_ijoin(tp, dp, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1230 unlock_dp_on_error = false;
1231
1232 error = xfs_dir_createname(tp, dp, name, ip->i_ino,
1233 &first_block, &free_list, resblks ?
1234 resblks - XFS_IALLOC_SPACE_RES(mp) : 0);
1235 if (error) {
1236 ASSERT(error != -ENOSPC);
1237 goto out_trans_cancel;
1238 }
1239 xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1240 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1241
1242 if (is_dir) {
1243 error = xfs_dir_init(tp, ip, dp);
1244 if (error)
1245 goto out_bmap_cancel;
1246
1247 error = xfs_bumplink(tp, dp);
1248 if (error)
1249 goto out_bmap_cancel;
1250 }
1251
1252 /*
1253 * If this is a synchronous mount, make sure that the
1254 * create transaction goes to disk before returning to
1255 * the user.
1256 */
1257 if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1258 xfs_trans_set_sync(tp);
1259
1260 /*
1261 * Attach the dquot(s) to the inodes and modify them incore.
1262 * These ids of the inode couldn't have changed since the new
1263 * inode has been locked ever since it was created.
1264 */
1265 xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1266
1267 error = xfs_bmap_finish(&tp, &free_list, NULL);
1268 if (error)
1269 goto out_bmap_cancel;
1270
1271 error = xfs_trans_commit(tp);
1272 if (error)
1273 goto out_release_inode;
1274
1275 xfs_qm_dqrele(udqp);
1276 xfs_qm_dqrele(gdqp);
1277 xfs_qm_dqrele(pdqp);
1278
1279 *ipp = ip;
1280 return 0;
1281
1282 out_bmap_cancel:
1283 xfs_bmap_cancel(&free_list);
1284 out_trans_cancel:
1285 xfs_trans_cancel(tp);
1286 out_release_inode:
1287 /*
1288 * Wait until after the current transaction is aborted to finish the
1289 * setup of the inode and release the inode. This prevents recursive
1290 * transactions and deadlocks from xfs_inactive.
1291 */
1292 if (ip) {
1293 xfs_finish_inode_setup(ip);
1294 IRELE(ip);
1295 }
1296
1297 xfs_qm_dqrele(udqp);
1298 xfs_qm_dqrele(gdqp);
1299 xfs_qm_dqrele(pdqp);
1300
1301 if (unlock_dp_on_error)
1302 xfs_iunlock(dp, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1303 return error;
1304}
1305
1306int
1307xfs_create_tmpfile(
1308 struct xfs_inode *dp,
1309 struct dentry *dentry,
1310 umode_t mode,
1311 struct xfs_inode **ipp)
1312{
1313 struct xfs_mount *mp = dp->i_mount;
1314 struct xfs_inode *ip = NULL;
1315 struct xfs_trans *tp = NULL;
1316 int error;
1317 prid_t prid;
1318 struct xfs_dquot *udqp = NULL;
1319 struct xfs_dquot *gdqp = NULL;
1320 struct xfs_dquot *pdqp = NULL;
1321 struct xfs_trans_res *tres;
1322 uint resblks;
1323
1324 if (XFS_FORCED_SHUTDOWN(mp))
1325 return -EIO;
1326
1327 prid = xfs_get_initial_prid(dp);
1328
1329 /*
1330 * Make sure that we have allocated dquot(s) on disk.
1331 */
1332 error = xfs_qm_vop_dqalloc(dp, xfs_kuid_to_uid(current_fsuid()),
1333 xfs_kgid_to_gid(current_fsgid()), prid,
1334 XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
1335 &udqp, &gdqp, &pdqp);
1336 if (error)
1337 return error;
1338
1339 resblks = XFS_IALLOC_SPACE_RES(mp);
1340 tp = xfs_trans_alloc(mp, XFS_TRANS_CREATE_TMPFILE);
1341
1342 tres = &M_RES(mp)->tr_create_tmpfile;
1343 error = xfs_trans_reserve(tp, tres, resblks, 0);
1344 if (error == -ENOSPC) {
1345 /* No space at all so try a "no-allocation" reservation */
1346 resblks = 0;
1347 error = xfs_trans_reserve(tp, tres, 0, 0);
1348 }
1349 if (error)
1350 goto out_trans_cancel;
1351
1352 error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp,
1353 pdqp, resblks, 1, 0);
1354 if (error)
1355 goto out_trans_cancel;
1356
1357 error = xfs_dir_ialloc(&tp, dp, mode, 1, 0,
1358 prid, resblks > 0, &ip, NULL);
1359 if (error)
1360 goto out_trans_cancel;
1361
1362 if (mp->m_flags & XFS_MOUNT_WSYNC)
1363 xfs_trans_set_sync(tp);
1364
1365 /*
1366 * Attach the dquot(s) to the inodes and modify them incore.
1367 * These ids of the inode couldn't have changed since the new
1368 * inode has been locked ever since it was created.
1369 */
1370 xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1371
1372 error = xfs_iunlink(tp, ip);
1373 if (error)
1374 goto out_trans_cancel;
1375
1376 error = xfs_trans_commit(tp);
1377 if (error)
1378 goto out_release_inode;
1379
1380 xfs_qm_dqrele(udqp);
1381 xfs_qm_dqrele(gdqp);
1382 xfs_qm_dqrele(pdqp);
1383
1384 *ipp = ip;
1385 return 0;
1386
1387 out_trans_cancel:
1388 xfs_trans_cancel(tp);
1389 out_release_inode:
1390 /*
1391 * Wait until after the current transaction is aborted to finish the
1392 * setup of the inode and release the inode. This prevents recursive
1393 * transactions and deadlocks from xfs_inactive.
1394 */
1395 if (ip) {
1396 xfs_finish_inode_setup(ip);
1397 IRELE(ip);
1398 }
1399
1400 xfs_qm_dqrele(udqp);
1401 xfs_qm_dqrele(gdqp);
1402 xfs_qm_dqrele(pdqp);
1403
1404 return error;
1405}
1406
1407int
1408xfs_link(
1409 xfs_inode_t *tdp,
1410 xfs_inode_t *sip,
1411 struct xfs_name *target_name)
1412{
1413 xfs_mount_t *mp = tdp->i_mount;
1414 xfs_trans_t *tp;
1415 int error;
1416 xfs_bmap_free_t free_list;
1417 xfs_fsblock_t first_block;
1418 int resblks;
1419
1420 trace_xfs_link(tdp, target_name);
1421
1422 ASSERT(!S_ISDIR(VFS_I(sip)->i_mode));
1423
1424 if (XFS_FORCED_SHUTDOWN(mp))
1425 return -EIO;
1426
1427 error = xfs_qm_dqattach(sip, 0);
1428 if (error)
1429 goto std_return;
1430
1431 error = xfs_qm_dqattach(tdp, 0);
1432 if (error)
1433 goto std_return;
1434
1435 tp = xfs_trans_alloc(mp, XFS_TRANS_LINK);
1436 resblks = XFS_LINK_SPACE_RES(mp, target_name->len);
1437 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_link, resblks, 0);
1438 if (error == -ENOSPC) {
1439 resblks = 0;
1440 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_link, 0, 0);
1441 }
1442 if (error)
1443 goto error_return;
1444
1445 xfs_ilock(tdp, XFS_IOLOCK_EXCL | XFS_IOLOCK_PARENT);
1446 xfs_lock_two_inodes(sip, tdp, XFS_ILOCK_EXCL);
1447
1448 xfs_trans_ijoin(tp, sip, XFS_ILOCK_EXCL);
1449 xfs_trans_ijoin(tp, tdp, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
1450
1451 /*
1452 * If we are using project inheritance, we only allow hard link
1453 * creation in our tree when the project IDs are the same; else
1454 * the tree quota mechanism could be circumvented.
1455 */
1456 if (unlikely((tdp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
1457 (xfs_get_projid(tdp) != xfs_get_projid(sip)))) {
1458 error = -EXDEV;
1459 goto error_return;
1460 }
1461
1462 if (!resblks) {
1463 error = xfs_dir_canenter(tp, tdp, target_name);
1464 if (error)
1465 goto error_return;
1466 }
1467
1468 xfs_bmap_init(&free_list, &first_block);
1469
1470 /*
1471 * Handle initial link state of O_TMPFILE inode
1472 */
1473 if (VFS_I(sip)->i_nlink == 0) {
1474 error = xfs_iunlink_remove(tp, sip);
1475 if (error)
1476 goto error_return;
1477 }
1478
1479 error = xfs_dir_createname(tp, tdp, target_name, sip->i_ino,
1480 &first_block, &free_list, resblks);
1481 if (error)
1482 goto error_return;
1483 xfs_trans_ichgtime(tp, tdp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1484 xfs_trans_log_inode(tp, tdp, XFS_ILOG_CORE);
1485
1486 error = xfs_bumplink(tp, sip);
1487 if (error)
1488 goto error_return;
1489
1490 /*
1491 * If this is a synchronous mount, make sure that the
1492 * link transaction goes to disk before returning to
1493 * the user.
1494 */
1495 if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1496 xfs_trans_set_sync(tp);
1497
1498 error = xfs_bmap_finish(&tp, &free_list, NULL);
1499 if (error) {
1500 xfs_bmap_cancel(&free_list);
1501 goto error_return;
1502 }
1503
1504 return xfs_trans_commit(tp);
1505
1506 error_return:
1507 xfs_trans_cancel(tp);
1508 std_return:
1509 return error;
1510}
1511
1512/*
1513 * Free up the underlying blocks past new_size. The new size must be smaller
1514 * than the current size. This routine can be used both for the attribute and
1515 * data fork, and does not modify the inode size, which is left to the caller.
1516 *
1517 * The transaction passed to this routine must have made a permanent log
1518 * reservation of at least XFS_ITRUNCATE_LOG_RES. This routine may commit the
1519 * given transaction and start new ones, so make sure everything involved in
1520 * the transaction is tidy before calling here. Some transaction will be
1521 * returned to the caller to be committed. The incoming transaction must
1522 * already include the inode, and both inode locks must be held exclusively.
1523 * The inode must also be "held" within the transaction. On return the inode
1524 * will be "held" within the returned transaction. This routine does NOT
1525 * require any disk space to be reserved for it within the transaction.
1526 *
1527 * If we get an error, we must return with the inode locked and linked into the
1528 * current transaction. This keeps things simple for the higher level code,
1529 * because it always knows that the inode is locked and held in the transaction
1530 * that returns to it whether errors occur or not. We don't mark the inode
1531 * dirty on error so that transactions can be easily aborted if possible.
1532 */
1533int
1534xfs_itruncate_extents(
1535 struct xfs_trans **tpp,
1536 struct xfs_inode *ip,
1537 int whichfork,
1538 xfs_fsize_t new_size)
1539{
1540 struct xfs_mount *mp = ip->i_mount;
1541 struct xfs_trans *tp = *tpp;
1542 xfs_bmap_free_t free_list;
1543 xfs_fsblock_t first_block;
1544 xfs_fileoff_t first_unmap_block;
1545 xfs_fileoff_t last_block;
1546 xfs_filblks_t unmap_len;
1547 int error = 0;
1548 int done = 0;
1549
1550 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1551 ASSERT(!atomic_read(&VFS_I(ip)->i_count) ||
1552 xfs_isilocked(ip, XFS_IOLOCK_EXCL));
1553 ASSERT(new_size <= XFS_ISIZE(ip));
1554 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
1555 ASSERT(ip->i_itemp != NULL);
1556 ASSERT(ip->i_itemp->ili_lock_flags == 0);
1557 ASSERT(!XFS_NOT_DQATTACHED(mp, ip));
1558
1559 trace_xfs_itruncate_extents_start(ip, new_size);
1560
1561 /*
1562 * Since it is possible for space to become allocated beyond
1563 * the end of the file (in a crash where the space is allocated
1564 * but the inode size is not yet updated), simply remove any
1565 * blocks which show up between the new EOF and the maximum
1566 * possible file size. If the first block to be removed is
1567 * beyond the maximum file size (ie it is the same as last_block),
1568 * then there is nothing to do.
1569 */
1570 first_unmap_block = XFS_B_TO_FSB(mp, (xfs_ufsize_t)new_size);
1571 last_block = XFS_B_TO_FSB(mp, mp->m_super->s_maxbytes);
1572 if (first_unmap_block == last_block)
1573 return 0;
1574
1575 ASSERT(first_unmap_block < last_block);
1576 unmap_len = last_block - first_unmap_block + 1;
1577 while (!done) {
1578 xfs_bmap_init(&free_list, &first_block);
1579 error = xfs_bunmapi(tp, ip,
1580 first_unmap_block, unmap_len,
1581 xfs_bmapi_aflag(whichfork),
1582 XFS_ITRUNC_MAX_EXTENTS,
1583 &first_block, &free_list,
1584 &done);
1585 if (error)
1586 goto out_bmap_cancel;
1587
1588 /*
1589 * Duplicate the transaction that has the permanent
1590 * reservation and commit the old transaction.
1591 */
1592 error = xfs_bmap_finish(&tp, &free_list, ip);
1593 if (error)
1594 goto out_bmap_cancel;
1595
1596 error = xfs_trans_roll(&tp, ip);
1597 if (error)
1598 goto out;
1599 }
1600
1601 /*
1602 * Always re-log the inode so that our permanent transaction can keep
1603 * on rolling it forward in the log.
1604 */
1605 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1606
1607 trace_xfs_itruncate_extents_end(ip, new_size);
1608
1609out:
1610 *tpp = tp;
1611 return error;
1612out_bmap_cancel:
1613 /*
1614 * If the bunmapi call encounters an error, return to the caller where
1615 * the transaction can be properly aborted. We just need to make sure
1616 * we're not holding any resources that we were not when we came in.
1617 */
1618 xfs_bmap_cancel(&free_list);
1619 goto out;
1620}
1621
1622int
1623xfs_release(
1624 xfs_inode_t *ip)
1625{
1626 xfs_mount_t *mp = ip->i_mount;
1627 int error;
1628
1629 if (!S_ISREG(VFS_I(ip)->i_mode) || (VFS_I(ip)->i_mode == 0))
1630 return 0;
1631
1632 /* If this is a read-only mount, don't do this (would generate I/O) */
1633 if (mp->m_flags & XFS_MOUNT_RDONLY)
1634 return 0;
1635
1636 if (!XFS_FORCED_SHUTDOWN(mp)) {
1637 int truncated;
1638
1639 /*
1640 * If we previously truncated this file and removed old data
1641 * in the process, we want to initiate "early" writeout on
1642 * the last close. This is an attempt to combat the notorious
1643 * NULL files problem which is particularly noticeable from a
1644 * truncate down, buffered (re-)write (delalloc), followed by
1645 * a crash. What we are effectively doing here is
1646 * significantly reducing the time window where we'd otherwise
1647 * be exposed to that problem.
1648 */
1649 truncated = xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED);
1650 if (truncated) {
1651 xfs_iflags_clear(ip, XFS_IDIRTY_RELEASE);
1652 if (ip->i_delayed_blks > 0) {
1653 error = filemap_flush(VFS_I(ip)->i_mapping);
1654 if (error)
1655 return error;
1656 }
1657 }
1658 }
1659
1660 if (VFS_I(ip)->i_nlink == 0)
1661 return 0;
1662
1663 if (xfs_can_free_eofblocks(ip, false)) {
1664
1665 /*
1666 * If we can't get the iolock just skip truncating the blocks
1667 * past EOF because we could deadlock with the mmap_sem
1668 * otherwise. We'll get another chance to drop them once the
1669 * last reference to the inode is dropped, so we'll never leak
1670 * blocks permanently.
1671 *
1672 * Further, check if the inode is being opened, written and
1673 * closed frequently and we have delayed allocation blocks
1674 * outstanding (e.g. streaming writes from the NFS server),
1675 * truncating the blocks past EOF will cause fragmentation to
1676 * occur.
1677 *
1678 * In this case don't do the truncation, either, but we have to
1679 * be careful how we detect this case. Blocks beyond EOF show
1680 * up as i_delayed_blks even when the inode is clean, so we
1681 * need to truncate them away first before checking for a dirty
1682 * release. Hence on the first dirty close we will still remove
1683 * the speculative allocation, but after that we will leave it
1684 * in place.
1685 */
1686 if (xfs_iflags_test(ip, XFS_IDIRTY_RELEASE))
1687 return 0;
1688
1689 error = xfs_free_eofblocks(mp, ip, true);
1690 if (error && error != -EAGAIN)
1691 return error;
1692
1693 /* delalloc blocks after truncation means it really is dirty */
1694 if (ip->i_delayed_blks)
1695 xfs_iflags_set(ip, XFS_IDIRTY_RELEASE);
1696 }
1697 return 0;
1698}
1699
1700/*
1701 * xfs_inactive_truncate
1702 *
1703 * Called to perform a truncate when an inode becomes unlinked.
1704 */
1705STATIC int
1706xfs_inactive_truncate(
1707 struct xfs_inode *ip)
1708{
1709 struct xfs_mount *mp = ip->i_mount;
1710 struct xfs_trans *tp;
1711 int error;
1712
1713 tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
1714 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_itruncate, 0, 0);
1715 if (error) {
1716 ASSERT(XFS_FORCED_SHUTDOWN(mp));
1717 xfs_trans_cancel(tp);
1718 return error;
1719 }
1720
1721 xfs_ilock(ip, XFS_ILOCK_EXCL);
1722 xfs_trans_ijoin(tp, ip, 0);
1723
1724 /*
1725 * Log the inode size first to prevent stale data exposure in the event
1726 * of a system crash before the truncate completes. See the related
1727 * comment in xfs_setattr_size() for details.
1728 */
1729 ip->i_d.di_size = 0;
1730 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1731
1732 error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK, 0);
1733 if (error)
1734 goto error_trans_cancel;
1735
1736 ASSERT(ip->i_d.di_nextents == 0);
1737
1738 error = xfs_trans_commit(tp);
1739 if (error)
1740 goto error_unlock;
1741
1742 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1743 return 0;
1744
1745error_trans_cancel:
1746 xfs_trans_cancel(tp);
1747error_unlock:
1748 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1749 return error;
1750}
1751
1752/*
1753 * xfs_inactive_ifree()
1754 *
1755 * Perform the inode free when an inode is unlinked.
1756 */
1757STATIC int
1758xfs_inactive_ifree(
1759 struct xfs_inode *ip)
1760{
1761 xfs_bmap_free_t free_list;
1762 xfs_fsblock_t first_block;
1763 struct xfs_mount *mp = ip->i_mount;
1764 struct xfs_trans *tp;
1765 int error;
1766
1767 tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
1768
1769 /*
1770 * The ifree transaction might need to allocate blocks for record
1771 * insertion to the finobt. We don't want to fail here at ENOSPC, so
1772 * allow ifree to dip into the reserved block pool if necessary.
1773 *
1774 * Freeing large sets of inodes generally means freeing inode chunks,
1775 * directory and file data blocks, so this should be relatively safe.
1776 * Only under severe circumstances should it be possible to free enough
1777 * inodes to exhaust the reserve block pool via finobt expansion while
1778 * at the same time not creating free space in the filesystem.
1779 *
1780 * Send a warning if the reservation does happen to fail, as the inode
1781 * now remains allocated and sits on the unlinked list until the fs is
1782 * repaired.
1783 */
1784 tp->t_flags |= XFS_TRANS_RESERVE;
1785 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_ifree,
1786 XFS_IFREE_SPACE_RES(mp), 0);
1787 if (error) {
1788 if (error == -ENOSPC) {
1789 xfs_warn_ratelimited(mp,
1790 "Failed to remove inode(s) from unlinked list. "
1791 "Please free space, unmount and run xfs_repair.");
1792 } else {
1793 ASSERT(XFS_FORCED_SHUTDOWN(mp));
1794 }
1795 xfs_trans_cancel(tp);
1796 return error;
1797 }
1798
1799 xfs_ilock(ip, XFS_ILOCK_EXCL);
1800 xfs_trans_ijoin(tp, ip, 0);
1801
1802 xfs_bmap_init(&free_list, &first_block);
1803 error = xfs_ifree(tp, ip, &free_list);
1804 if (error) {
1805 /*
1806 * If we fail to free the inode, shut down. The cancel
1807 * might do that, we need to make sure. Otherwise the
1808 * inode might be lost for a long time or forever.
1809 */
1810 if (!XFS_FORCED_SHUTDOWN(mp)) {
1811 xfs_notice(mp, "%s: xfs_ifree returned error %d",
1812 __func__, error);
1813 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1814 }
1815 xfs_trans_cancel(tp);
1816 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1817 return error;
1818 }
1819
1820 /*
1821 * Credit the quota account(s). The inode is gone.
1822 */
1823 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
1824
1825 /*
1826 * Just ignore errors at this point. There is nothing we can do except
1827 * to try to keep going. Make sure it's not a silent error.
1828 */
1829 error = xfs_bmap_finish(&tp, &free_list, NULL);
1830 if (error) {
1831 xfs_notice(mp, "%s: xfs_bmap_finish returned error %d",
1832 __func__, error);
1833 xfs_bmap_cancel(&free_list);
1834 }
1835 error = xfs_trans_commit(tp);
1836 if (error)
1837 xfs_notice(mp, "%s: xfs_trans_commit returned error %d",
1838 __func__, error);
1839
1840 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1841 return 0;
1842}
1843
1844/*
1845 * xfs_inactive
1846 *
1847 * This is called when the vnode reference count for the vnode
1848 * goes to zero. If the file has been unlinked, then it must
1849 * now be truncated. Also, we clear all of the read-ahead state
1850 * kept for the inode here since the file is now closed.
1851 */
1852void
1853xfs_inactive(
1854 xfs_inode_t *ip)
1855{
1856 struct xfs_mount *mp;
1857 int error;
1858 int truncate = 0;
1859
1860 /*
1861 * If the inode is already free, then there can be nothing
1862 * to clean up here.
1863 */
1864 if (VFS_I(ip)->i_mode == 0) {
1865 ASSERT(ip->i_df.if_real_bytes == 0);
1866 ASSERT(ip->i_df.if_broot_bytes == 0);
1867 return;
1868 }
1869
1870 mp = ip->i_mount;
1871
1872 /* If this is a read-only mount, don't do this (would generate I/O) */
1873 if (mp->m_flags & XFS_MOUNT_RDONLY)
1874 return;
1875
1876 if (VFS_I(ip)->i_nlink != 0) {
1877 /*
1878 * force is true because we are evicting an inode from the
1879 * cache. Post-eof blocks must be freed, lest we end up with
1880 * broken free space accounting.
1881 */
1882 if (xfs_can_free_eofblocks(ip, true))
1883 xfs_free_eofblocks(mp, ip, false);
1884
1885 return;
1886 }
1887
1888 if (S_ISREG(VFS_I(ip)->i_mode) &&
1889 (ip->i_d.di_size != 0 || XFS_ISIZE(ip) != 0 ||
1890 ip->i_d.di_nextents > 0 || ip->i_delayed_blks > 0))
1891 truncate = 1;
1892
1893 error = xfs_qm_dqattach(ip, 0);
1894 if (error)
1895 return;
1896
1897 if (S_ISLNK(VFS_I(ip)->i_mode))
1898 error = xfs_inactive_symlink(ip);
1899 else if (truncate)
1900 error = xfs_inactive_truncate(ip);
1901 if (error)
1902 return;
1903
1904 /*
1905 * If there are attributes associated with the file then blow them away
1906 * now. The code calls a routine that recursively deconstructs the
1907 * attribute fork. If also blows away the in-core attribute fork.
1908 */
1909 if (XFS_IFORK_Q(ip)) {
1910 error = xfs_attr_inactive(ip);
1911 if (error)
1912 return;
1913 }
1914
1915 ASSERT(!ip->i_afp);
1916 ASSERT(ip->i_d.di_anextents == 0);
1917 ASSERT(ip->i_d.di_forkoff == 0);
1918
1919 /*
1920 * Free the inode.
1921 */
1922 error = xfs_inactive_ifree(ip);
1923 if (error)
1924 return;
1925
1926 /*
1927 * Release the dquots held by inode, if any.
1928 */
1929 xfs_qm_dqdetach(ip);
1930}
1931
1932/*
1933 * This is called when the inode's link count goes to 0 or we are creating a
1934 * tmpfile via O_TMPFILE. In the case of a tmpfile, @ignore_linkcount will be
1935 * set to true as the link count is dropped to zero by the VFS after we've
1936 * created the file successfully, so we have to add it to the unlinked list
1937 * while the link count is non-zero.
1938 *
1939 * We place the on-disk inode on a list in the AGI. It will be pulled from this
1940 * list when the inode is freed.
1941 */
1942STATIC int
1943xfs_iunlink(
1944 struct xfs_trans *tp,
1945 struct xfs_inode *ip)
1946{
1947 xfs_mount_t *mp = tp->t_mountp;
1948 xfs_agi_t *agi;
1949 xfs_dinode_t *dip;
1950 xfs_buf_t *agibp;
1951 xfs_buf_t *ibp;
1952 xfs_agino_t agino;
1953 short bucket_index;
1954 int offset;
1955 int error;
1956
1957 ASSERT(VFS_I(ip)->i_mode != 0);
1958
1959 /*
1960 * Get the agi buffer first. It ensures lock ordering
1961 * on the list.
1962 */
1963 error = xfs_read_agi(mp, tp, XFS_INO_TO_AGNO(mp, ip->i_ino), &agibp);
1964 if (error)
1965 return error;
1966 agi = XFS_BUF_TO_AGI(agibp);
1967
1968 /*
1969 * Get the index into the agi hash table for the
1970 * list this inode will go on.
1971 */
1972 agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
1973 ASSERT(agino != 0);
1974 bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
1975 ASSERT(agi->agi_unlinked[bucket_index]);
1976 ASSERT(be32_to_cpu(agi->agi_unlinked[bucket_index]) != agino);
1977
1978 if (agi->agi_unlinked[bucket_index] != cpu_to_be32(NULLAGINO)) {
1979 /*
1980 * There is already another inode in the bucket we need
1981 * to add ourselves to. Add us at the front of the list.
1982 * Here we put the head pointer into our next pointer,
1983 * and then we fall through to point the head at us.
1984 */
1985 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &dip, &ibp,
1986 0, 0);
1987 if (error)
1988 return error;
1989
1990 ASSERT(dip->di_next_unlinked == cpu_to_be32(NULLAGINO));
1991 dip->di_next_unlinked = agi->agi_unlinked[bucket_index];
1992 offset = ip->i_imap.im_boffset +
1993 offsetof(xfs_dinode_t, di_next_unlinked);
1994
1995 /* need to recalc the inode CRC if appropriate */
1996 xfs_dinode_calc_crc(mp, dip);
1997
1998 xfs_trans_inode_buf(tp, ibp);
1999 xfs_trans_log_buf(tp, ibp, offset,
2000 (offset + sizeof(xfs_agino_t) - 1));
2001 xfs_inobp_check(mp, ibp);
2002 }
2003
2004 /*
2005 * Point the bucket head pointer at the inode being inserted.
2006 */
2007 ASSERT(agino != 0);
2008 agi->agi_unlinked[bucket_index] = cpu_to_be32(agino);
2009 offset = offsetof(xfs_agi_t, agi_unlinked) +
2010 (sizeof(xfs_agino_t) * bucket_index);
2011 xfs_trans_buf_set_type(tp, agibp, XFS_BLFT_AGI_BUF);
2012 xfs_trans_log_buf(tp, agibp, offset,
2013 (offset + sizeof(xfs_agino_t) - 1));
2014 return 0;
2015}
2016
2017/*
2018 * Pull the on-disk inode from the AGI unlinked list.
2019 */
2020STATIC int
2021xfs_iunlink_remove(
2022 xfs_trans_t *tp,
2023 xfs_inode_t *ip)
2024{
2025 xfs_ino_t next_ino;
2026 xfs_mount_t *mp;
2027 xfs_agi_t *agi;
2028 xfs_dinode_t *dip;
2029 xfs_buf_t *agibp;
2030 xfs_buf_t *ibp;
2031 xfs_agnumber_t agno;
2032 xfs_agino_t agino;
2033 xfs_agino_t next_agino;
2034 xfs_buf_t *last_ibp;
2035 xfs_dinode_t *last_dip = NULL;
2036 short bucket_index;
2037 int offset, last_offset = 0;
2038 int error;
2039
2040 mp = tp->t_mountp;
2041 agno = XFS_INO_TO_AGNO(mp, ip->i_ino);
2042
2043 /*
2044 * Get the agi buffer first. It ensures lock ordering
2045 * on the list.
2046 */
2047 error = xfs_read_agi(mp, tp, agno, &agibp);
2048 if (error)
2049 return error;
2050
2051 agi = XFS_BUF_TO_AGI(agibp);
2052
2053 /*
2054 * Get the index into the agi hash table for the
2055 * list this inode will go on.
2056 */
2057 agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2058 ASSERT(agino != 0);
2059 bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
2060 ASSERT(agi->agi_unlinked[bucket_index] != cpu_to_be32(NULLAGINO));
2061 ASSERT(agi->agi_unlinked[bucket_index]);
2062
2063 if (be32_to_cpu(agi->agi_unlinked[bucket_index]) == agino) {
2064 /*
2065 * We're at the head of the list. Get the inode's on-disk
2066 * buffer to see if there is anyone after us on the list.
2067 * Only modify our next pointer if it is not already NULLAGINO.
2068 * This saves us the overhead of dealing with the buffer when
2069 * there is no need to change it.
2070 */
2071 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &dip, &ibp,
2072 0, 0);
2073 if (error) {
2074 xfs_warn(mp, "%s: xfs_imap_to_bp returned error %d.",
2075 __func__, error);
2076 return error;
2077 }
2078 next_agino = be32_to_cpu(dip->di_next_unlinked);
2079 ASSERT(next_agino != 0);
2080 if (next_agino != NULLAGINO) {
2081 dip->di_next_unlinked = cpu_to_be32(NULLAGINO);
2082 offset = ip->i_imap.im_boffset +
2083 offsetof(xfs_dinode_t, di_next_unlinked);
2084
2085 /* need to recalc the inode CRC if appropriate */
2086 xfs_dinode_calc_crc(mp, dip);
2087
2088 xfs_trans_inode_buf(tp, ibp);
2089 xfs_trans_log_buf(tp, ibp, offset,
2090 (offset + sizeof(xfs_agino_t) - 1));
2091 xfs_inobp_check(mp, ibp);
2092 } else {
2093 xfs_trans_brelse(tp, ibp);
2094 }
2095 /*
2096 * Point the bucket head pointer at the next inode.
2097 */
2098 ASSERT(next_agino != 0);
2099 ASSERT(next_agino != agino);
2100 agi->agi_unlinked[bucket_index] = cpu_to_be32(next_agino);
2101 offset = offsetof(xfs_agi_t, agi_unlinked) +
2102 (sizeof(xfs_agino_t) * bucket_index);
2103 xfs_trans_buf_set_type(tp, agibp, XFS_BLFT_AGI_BUF);
2104 xfs_trans_log_buf(tp, agibp, offset,
2105 (offset + sizeof(xfs_agino_t) - 1));
2106 } else {
2107 /*
2108 * We need to search the list for the inode being freed.
2109 */
2110 next_agino = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2111 last_ibp = NULL;
2112 while (next_agino != agino) {
2113 struct xfs_imap imap;
2114
2115 if (last_ibp)
2116 xfs_trans_brelse(tp, last_ibp);
2117
2118 imap.im_blkno = 0;
2119 next_ino = XFS_AGINO_TO_INO(mp, agno, next_agino);
2120
2121 error = xfs_imap(mp, tp, next_ino, &imap, 0);
2122 if (error) {
2123 xfs_warn(mp,
2124 "%s: xfs_imap returned error %d.",
2125 __func__, error);
2126 return error;
2127 }
2128
2129 error = xfs_imap_to_bp(mp, tp, &imap, &last_dip,
2130 &last_ibp, 0, 0);
2131 if (error) {
2132 xfs_warn(mp,
2133 "%s: xfs_imap_to_bp returned error %d.",
2134 __func__, error);
2135 return error;
2136 }
2137
2138 last_offset = imap.im_boffset;
2139 next_agino = be32_to_cpu(last_dip->di_next_unlinked);
2140 ASSERT(next_agino != NULLAGINO);
2141 ASSERT(next_agino != 0);
2142 }
2143
2144 /*
2145 * Now last_ibp points to the buffer previous to us on the
2146 * unlinked list. Pull us from the list.
2147 */
2148 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &dip, &ibp,
2149 0, 0);
2150 if (error) {
2151 xfs_warn(mp, "%s: xfs_imap_to_bp(2) returned error %d.",
2152 __func__, error);
2153 return error;
2154 }
2155 next_agino = be32_to_cpu(dip->di_next_unlinked);
2156 ASSERT(next_agino != 0);
2157 ASSERT(next_agino != agino);
2158 if (next_agino != NULLAGINO) {
2159 dip->di_next_unlinked = cpu_to_be32(NULLAGINO);
2160 offset = ip->i_imap.im_boffset +
2161 offsetof(xfs_dinode_t, di_next_unlinked);
2162
2163 /* need to recalc the inode CRC if appropriate */
2164 xfs_dinode_calc_crc(mp, dip);
2165
2166 xfs_trans_inode_buf(tp, ibp);
2167 xfs_trans_log_buf(tp, ibp, offset,
2168 (offset + sizeof(xfs_agino_t) - 1));
2169 xfs_inobp_check(mp, ibp);
2170 } else {
2171 xfs_trans_brelse(tp, ibp);
2172 }
2173 /*
2174 * Point the previous inode on the list to the next inode.
2175 */
2176 last_dip->di_next_unlinked = cpu_to_be32(next_agino);
2177 ASSERT(next_agino != 0);
2178 offset = last_offset + offsetof(xfs_dinode_t, di_next_unlinked);
2179
2180 /* need to recalc the inode CRC if appropriate */
2181 xfs_dinode_calc_crc(mp, last_dip);
2182
2183 xfs_trans_inode_buf(tp, last_ibp);
2184 xfs_trans_log_buf(tp, last_ibp, offset,
2185 (offset + sizeof(xfs_agino_t) - 1));
2186 xfs_inobp_check(mp, last_ibp);
2187 }
2188 return 0;
2189}
2190
2191/*
2192 * A big issue when freeing the inode cluster is that we _cannot_ skip any
2193 * inodes that are in memory - they all must be marked stale and attached to
2194 * the cluster buffer.
2195 */
2196STATIC int
2197xfs_ifree_cluster(
2198 xfs_inode_t *free_ip,
2199 xfs_trans_t *tp,
2200 struct xfs_icluster *xic)
2201{
2202 xfs_mount_t *mp = free_ip->i_mount;
2203 int blks_per_cluster;
2204 int inodes_per_cluster;
2205 int nbufs;
2206 int i, j;
2207 int ioffset;
2208 xfs_daddr_t blkno;
2209 xfs_buf_t *bp;
2210 xfs_inode_t *ip;
2211 xfs_inode_log_item_t *iip;
2212 xfs_log_item_t *lip;
2213 struct xfs_perag *pag;
2214 xfs_ino_t inum;
2215
2216 inum = xic->first_ino;
2217 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, inum));
2218 blks_per_cluster = xfs_icluster_size_fsb(mp);
2219 inodes_per_cluster = blks_per_cluster << mp->m_sb.sb_inopblog;
2220 nbufs = mp->m_ialloc_blks / blks_per_cluster;
2221
2222 for (j = 0; j < nbufs; j++, inum += inodes_per_cluster) {
2223 /*
2224 * The allocation bitmap tells us which inodes of the chunk were
2225 * physically allocated. Skip the cluster if an inode falls into
2226 * a sparse region.
2227 */
2228 ioffset = inum - xic->first_ino;
2229 if ((xic->alloc & XFS_INOBT_MASK(ioffset)) == 0) {
2230 ASSERT(do_mod(ioffset, inodes_per_cluster) == 0);
2231 continue;
2232 }
2233
2234 blkno = XFS_AGB_TO_DADDR(mp, XFS_INO_TO_AGNO(mp, inum),
2235 XFS_INO_TO_AGBNO(mp, inum));
2236
2237 /*
2238 * We obtain and lock the backing buffer first in the process
2239 * here, as we have to ensure that any dirty inode that we
2240 * can't get the flush lock on is attached to the buffer.
2241 * If we scan the in-memory inodes first, then buffer IO can
2242 * complete before we get a lock on it, and hence we may fail
2243 * to mark all the active inodes on the buffer stale.
2244 */
2245 bp = xfs_trans_get_buf(tp, mp->m_ddev_targp, blkno,
2246 mp->m_bsize * blks_per_cluster,
2247 XBF_UNMAPPED);
2248
2249 if (!bp)
2250 return -ENOMEM;
2251
2252 /*
2253 * This buffer may not have been correctly initialised as we
2254 * didn't read it from disk. That's not important because we are
2255 * only using to mark the buffer as stale in the log, and to
2256 * attach stale cached inodes on it. That means it will never be
2257 * dispatched for IO. If it is, we want to know about it, and we
2258 * want it to fail. We can acheive this by adding a write
2259 * verifier to the buffer.
2260 */
2261 bp->b_ops = &xfs_inode_buf_ops;
2262
2263 /*
2264 * Walk the inodes already attached to the buffer and mark them
2265 * stale. These will all have the flush locks held, so an
2266 * in-memory inode walk can't lock them. By marking them all
2267 * stale first, we will not attempt to lock them in the loop
2268 * below as the XFS_ISTALE flag will be set.
2269 */
2270 lip = bp->b_fspriv;
2271 while (lip) {
2272 if (lip->li_type == XFS_LI_INODE) {
2273 iip = (xfs_inode_log_item_t *)lip;
2274 ASSERT(iip->ili_logged == 1);
2275 lip->li_cb = xfs_istale_done;
2276 xfs_trans_ail_copy_lsn(mp->m_ail,
2277 &iip->ili_flush_lsn,
2278 &iip->ili_item.li_lsn);
2279 xfs_iflags_set(iip->ili_inode, XFS_ISTALE);
2280 }
2281 lip = lip->li_bio_list;
2282 }
2283
2284
2285 /*
2286 * For each inode in memory attempt to add it to the inode
2287 * buffer and set it up for being staled on buffer IO
2288 * completion. This is safe as we've locked out tail pushing
2289 * and flushing by locking the buffer.
2290 *
2291 * We have already marked every inode that was part of a
2292 * transaction stale above, which means there is no point in
2293 * even trying to lock them.
2294 */
2295 for (i = 0; i < inodes_per_cluster; i++) {
2296retry:
2297 rcu_read_lock();
2298 ip = radix_tree_lookup(&pag->pag_ici_root,
2299 XFS_INO_TO_AGINO(mp, (inum + i)));
2300
2301 /* Inode not in memory, nothing to do */
2302 if (!ip) {
2303 rcu_read_unlock();
2304 continue;
2305 }
2306
2307 /*
2308 * because this is an RCU protected lookup, we could
2309 * find a recently freed or even reallocated inode
2310 * during the lookup. We need to check under the
2311 * i_flags_lock for a valid inode here. Skip it if it
2312 * is not valid, the wrong inode or stale.
2313 */
2314 spin_lock(&ip->i_flags_lock);
2315 if (ip->i_ino != inum + i ||
2316 __xfs_iflags_test(ip, XFS_ISTALE)) {
2317 spin_unlock(&ip->i_flags_lock);
2318 rcu_read_unlock();
2319 continue;
2320 }
2321 spin_unlock(&ip->i_flags_lock);
2322
2323 /*
2324 * Don't try to lock/unlock the current inode, but we
2325 * _cannot_ skip the other inodes that we did not find
2326 * in the list attached to the buffer and are not
2327 * already marked stale. If we can't lock it, back off
2328 * and retry.
2329 */
2330 if (ip != free_ip &&
2331 !xfs_ilock_nowait(ip, XFS_ILOCK_EXCL)) {
2332 rcu_read_unlock();
2333 delay(1);
2334 goto retry;
2335 }
2336 rcu_read_unlock();
2337
2338 xfs_iflock(ip);
2339 xfs_iflags_set(ip, XFS_ISTALE);
2340
2341 /*
2342 * we don't need to attach clean inodes or those only
2343 * with unlogged changes (which we throw away, anyway).
2344 */
2345 iip = ip->i_itemp;
2346 if (!iip || xfs_inode_clean(ip)) {
2347 ASSERT(ip != free_ip);
2348 xfs_ifunlock(ip);
2349 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2350 continue;
2351 }
2352
2353 iip->ili_last_fields = iip->ili_fields;
2354 iip->ili_fields = 0;
2355 iip->ili_fsync_fields = 0;
2356 iip->ili_logged = 1;
2357 xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
2358 &iip->ili_item.li_lsn);
2359
2360 xfs_buf_attach_iodone(bp, xfs_istale_done,
2361 &iip->ili_item);
2362
2363 if (ip != free_ip)
2364 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2365 }
2366
2367 xfs_trans_stale_inode_buf(tp, bp);
2368 xfs_trans_binval(tp, bp);
2369 }
2370
2371 xfs_perag_put(pag);
2372 return 0;
2373}
2374
2375/*
2376 * This is called to return an inode to the inode free list.
2377 * The inode should already be truncated to 0 length and have
2378 * no pages associated with it. This routine also assumes that
2379 * the inode is already a part of the transaction.
2380 *
2381 * The on-disk copy of the inode will have been added to the list
2382 * of unlinked inodes in the AGI. We need to remove the inode from
2383 * that list atomically with respect to freeing it here.
2384 */
2385int
2386xfs_ifree(
2387 xfs_trans_t *tp,
2388 xfs_inode_t *ip,
2389 xfs_bmap_free_t *flist)
2390{
2391 int error;
2392 struct xfs_icluster xic = { 0 };
2393
2394 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
2395 ASSERT(VFS_I(ip)->i_nlink == 0);
2396 ASSERT(ip->i_d.di_nextents == 0);
2397 ASSERT(ip->i_d.di_anextents == 0);
2398 ASSERT(ip->i_d.di_size == 0 || !S_ISREG(VFS_I(ip)->i_mode));
2399 ASSERT(ip->i_d.di_nblocks == 0);
2400
2401 /*
2402 * Pull the on-disk inode from the AGI unlinked list.
2403 */
2404 error = xfs_iunlink_remove(tp, ip);
2405 if (error)
2406 return error;
2407
2408 error = xfs_difree(tp, ip->i_ino, flist, &xic);
2409 if (error)
2410 return error;
2411
2412 VFS_I(ip)->i_mode = 0; /* mark incore inode as free */
2413 ip->i_d.di_flags = 0;
2414 ip->i_d.di_dmevmask = 0;
2415 ip->i_d.di_forkoff = 0; /* mark the attr fork not in use */
2416 ip->i_d.di_format = XFS_DINODE_FMT_EXTENTS;
2417 ip->i_d.di_aformat = XFS_DINODE_FMT_EXTENTS;
2418 /*
2419 * Bump the generation count so no one will be confused
2420 * by reincarnations of this inode.
2421 */
2422 VFS_I(ip)->i_generation++;
2423 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2424
2425 if (xic.deleted)
2426 error = xfs_ifree_cluster(ip, tp, &xic);
2427
2428 return error;
2429}
2430
2431/*
2432 * This is called to unpin an inode. The caller must have the inode locked
2433 * in at least shared mode so that the buffer cannot be subsequently pinned
2434 * once someone is waiting for it to be unpinned.
2435 */
2436static void
2437xfs_iunpin(
2438 struct xfs_inode *ip)
2439{
2440 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
2441
2442 trace_xfs_inode_unpin_nowait(ip, _RET_IP_);
2443
2444 /* Give the log a push to start the unpinning I/O */
2445 xfs_log_force_lsn(ip->i_mount, ip->i_itemp->ili_last_lsn, 0);
2446
2447}
2448
2449static void
2450__xfs_iunpin_wait(
2451 struct xfs_inode *ip)
2452{
2453 wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_IPINNED_BIT);
2454 DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IPINNED_BIT);
2455
2456 xfs_iunpin(ip);
2457
2458 do {
2459 prepare_to_wait(wq, &wait.wait, TASK_UNINTERRUPTIBLE);
2460 if (xfs_ipincount(ip))
2461 io_schedule();
2462 } while (xfs_ipincount(ip));
2463 finish_wait(wq, &wait.wait);
2464}
2465
2466void
2467xfs_iunpin_wait(
2468 struct xfs_inode *ip)
2469{
2470 if (xfs_ipincount(ip))
2471 __xfs_iunpin_wait(ip);
2472}
2473
2474/*
2475 * Removing an inode from the namespace involves removing the directory entry
2476 * and dropping the link count on the inode. Removing the directory entry can
2477 * result in locking an AGF (directory blocks were freed) and removing a link
2478 * count can result in placing the inode on an unlinked list which results in
2479 * locking an AGI.
2480 *
2481 * The big problem here is that we have an ordering constraint on AGF and AGI
2482 * locking - inode allocation locks the AGI, then can allocate a new extent for
2483 * new inodes, locking the AGF after the AGI. Similarly, freeing the inode
2484 * removes the inode from the unlinked list, requiring that we lock the AGI
2485 * first, and then freeing the inode can result in an inode chunk being freed
2486 * and hence freeing disk space requiring that we lock an AGF.
2487 *
2488 * Hence the ordering that is imposed by other parts of the code is AGI before
2489 * AGF. This means we cannot remove the directory entry before we drop the inode
2490 * reference count and put it on the unlinked list as this results in a lock
2491 * order of AGF then AGI, and this can deadlock against inode allocation and
2492 * freeing. Therefore we must drop the link counts before we remove the
2493 * directory entry.
2494 *
2495 * This is still safe from a transactional point of view - it is not until we
2496 * get to xfs_bmap_finish() that we have the possibility of multiple
2497 * transactions in this operation. Hence as long as we remove the directory
2498 * entry and drop the link count in the first transaction of the remove
2499 * operation, there are no transactional constraints on the ordering here.
2500 */
2501int
2502xfs_remove(
2503 xfs_inode_t *dp,
2504 struct xfs_name *name,
2505 xfs_inode_t *ip)
2506{
2507 xfs_mount_t *mp = dp->i_mount;
2508 xfs_trans_t *tp = NULL;
2509 int is_dir = S_ISDIR(VFS_I(ip)->i_mode);
2510 int error = 0;
2511 xfs_bmap_free_t free_list;
2512 xfs_fsblock_t first_block;
2513 uint resblks;
2514
2515 trace_xfs_remove(dp, name);
2516
2517 if (XFS_FORCED_SHUTDOWN(mp))
2518 return -EIO;
2519
2520 error = xfs_qm_dqattach(dp, 0);
2521 if (error)
2522 goto std_return;
2523
2524 error = xfs_qm_dqattach(ip, 0);
2525 if (error)
2526 goto std_return;
2527
2528 if (is_dir)
2529 tp = xfs_trans_alloc(mp, XFS_TRANS_RMDIR);
2530 else
2531 tp = xfs_trans_alloc(mp, XFS_TRANS_REMOVE);
2532
2533 /*
2534 * We try to get the real space reservation first,
2535 * allowing for directory btree deletion(s) implying
2536 * possible bmap insert(s). If we can't get the space
2537 * reservation then we use 0 instead, and avoid the bmap
2538 * btree insert(s) in the directory code by, if the bmap
2539 * insert tries to happen, instead trimming the LAST
2540 * block from the directory.
2541 */
2542 resblks = XFS_REMOVE_SPACE_RES(mp);
2543 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_remove, resblks, 0);
2544 if (error == -ENOSPC) {
2545 resblks = 0;
2546 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_remove, 0, 0);
2547 }
2548 if (error) {
2549 ASSERT(error != -ENOSPC);
2550 goto out_trans_cancel;
2551 }
2552
2553 xfs_ilock(dp, XFS_IOLOCK_EXCL | XFS_IOLOCK_PARENT);
2554 xfs_lock_two_inodes(dp, ip, XFS_ILOCK_EXCL);
2555
2556 xfs_trans_ijoin(tp, dp, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
2557 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
2558
2559 /*
2560 * If we're removing a directory perform some additional validation.
2561 */
2562 if (is_dir) {
2563 ASSERT(VFS_I(ip)->i_nlink >= 2);
2564 if (VFS_I(ip)->i_nlink != 2) {
2565 error = -ENOTEMPTY;
2566 goto out_trans_cancel;
2567 }
2568 if (!xfs_dir_isempty(ip)) {
2569 error = -ENOTEMPTY;
2570 goto out_trans_cancel;
2571 }
2572
2573 /* Drop the link from ip's "..". */
2574 error = xfs_droplink(tp, dp);
2575 if (error)
2576 goto out_trans_cancel;
2577
2578 /* Drop the "." link from ip to self. */
2579 error = xfs_droplink(tp, ip);
2580 if (error)
2581 goto out_trans_cancel;
2582 } else {
2583 /*
2584 * When removing a non-directory we need to log the parent
2585 * inode here. For a directory this is done implicitly
2586 * by the xfs_droplink call for the ".." entry.
2587 */
2588 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
2589 }
2590 xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2591
2592 /* Drop the link from dp to ip. */
2593 error = xfs_droplink(tp, ip);
2594 if (error)
2595 goto out_trans_cancel;
2596
2597 xfs_bmap_init(&free_list, &first_block);
2598 error = xfs_dir_removename(tp, dp, name, ip->i_ino,
2599 &first_block, &free_list, resblks);
2600 if (error) {
2601 ASSERT(error != -ENOENT);
2602 goto out_bmap_cancel;
2603 }
2604
2605 /*
2606 * If this is a synchronous mount, make sure that the
2607 * remove transaction goes to disk before returning to
2608 * the user.
2609 */
2610 if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
2611 xfs_trans_set_sync(tp);
2612
2613 error = xfs_bmap_finish(&tp, &free_list, NULL);
2614 if (error)
2615 goto out_bmap_cancel;
2616
2617 error = xfs_trans_commit(tp);
2618 if (error)
2619 goto std_return;
2620
2621 if (is_dir && xfs_inode_is_filestream(ip))
2622 xfs_filestream_deassociate(ip);
2623
2624 return 0;
2625
2626 out_bmap_cancel:
2627 xfs_bmap_cancel(&free_list);
2628 out_trans_cancel:
2629 xfs_trans_cancel(tp);
2630 std_return:
2631 return error;
2632}
2633
2634/*
2635 * Enter all inodes for a rename transaction into a sorted array.
2636 */
2637#define __XFS_SORT_INODES 5
2638STATIC void
2639xfs_sort_for_rename(
2640 struct xfs_inode *dp1, /* in: old (source) directory inode */
2641 struct xfs_inode *dp2, /* in: new (target) directory inode */
2642 struct xfs_inode *ip1, /* in: inode of old entry */
2643 struct xfs_inode *ip2, /* in: inode of new entry */
2644 struct xfs_inode *wip, /* in: whiteout inode */
2645 struct xfs_inode **i_tab,/* out: sorted array of inodes */
2646 int *num_inodes) /* in/out: inodes in array */
2647{
2648 int i, j;
2649
2650 ASSERT(*num_inodes == __XFS_SORT_INODES);
2651 memset(i_tab, 0, *num_inodes * sizeof(struct xfs_inode *));
2652
2653 /*
2654 * i_tab contains a list of pointers to inodes. We initialize
2655 * the table here & we'll sort it. We will then use it to
2656 * order the acquisition of the inode locks.
2657 *
2658 * Note that the table may contain duplicates. e.g., dp1 == dp2.
2659 */
2660 i = 0;
2661 i_tab[i++] = dp1;
2662 i_tab[i++] = dp2;
2663 i_tab[i++] = ip1;
2664 if (ip2)
2665 i_tab[i++] = ip2;
2666 if (wip)
2667 i_tab[i++] = wip;
2668 *num_inodes = i;
2669
2670 /*
2671 * Sort the elements via bubble sort. (Remember, there are at
2672 * most 5 elements to sort, so this is adequate.)
2673 */
2674 for (i = 0; i < *num_inodes; i++) {
2675 for (j = 1; j < *num_inodes; j++) {
2676 if (i_tab[j]->i_ino < i_tab[j-1]->i_ino) {
2677 struct xfs_inode *temp = i_tab[j];
2678 i_tab[j] = i_tab[j-1];
2679 i_tab[j-1] = temp;
2680 }
2681 }
2682 }
2683}
2684
2685static int
2686xfs_finish_rename(
2687 struct xfs_trans *tp,
2688 struct xfs_bmap_free *free_list)
2689{
2690 int error;
2691
2692 /*
2693 * If this is a synchronous mount, make sure that the rename transaction
2694 * goes to disk before returning to the user.
2695 */
2696 if (tp->t_mountp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
2697 xfs_trans_set_sync(tp);
2698
2699 error = xfs_bmap_finish(&tp, free_list, NULL);
2700 if (error) {
2701 xfs_bmap_cancel(free_list);
2702 xfs_trans_cancel(tp);
2703 return error;
2704 }
2705
2706 return xfs_trans_commit(tp);
2707}
2708
2709/*
2710 * xfs_cross_rename()
2711 *
2712 * responsible for handling RENAME_EXCHANGE flag in renameat2() sytemcall
2713 */
2714STATIC int
2715xfs_cross_rename(
2716 struct xfs_trans *tp,
2717 struct xfs_inode *dp1,
2718 struct xfs_name *name1,
2719 struct xfs_inode *ip1,
2720 struct xfs_inode *dp2,
2721 struct xfs_name *name2,
2722 struct xfs_inode *ip2,
2723 struct xfs_bmap_free *free_list,
2724 xfs_fsblock_t *first_block,
2725 int spaceres)
2726{
2727 int error = 0;
2728 int ip1_flags = 0;
2729 int ip2_flags = 0;
2730 int dp2_flags = 0;
2731
2732 /* Swap inode number for dirent in first parent */
2733 error = xfs_dir_replace(tp, dp1, name1,
2734 ip2->i_ino,
2735 first_block, free_list, spaceres);
2736 if (error)
2737 goto out_trans_abort;
2738
2739 /* Swap inode number for dirent in second parent */
2740 error = xfs_dir_replace(tp, dp2, name2,
2741 ip1->i_ino,
2742 first_block, free_list, spaceres);
2743 if (error)
2744 goto out_trans_abort;
2745
2746 /*
2747 * If we're renaming one or more directories across different parents,
2748 * update the respective ".." entries (and link counts) to match the new
2749 * parents.
2750 */
2751 if (dp1 != dp2) {
2752 dp2_flags = XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2753
2754 if (S_ISDIR(VFS_I(ip2)->i_mode)) {
2755 error = xfs_dir_replace(tp, ip2, &xfs_name_dotdot,
2756 dp1->i_ino, first_block,
2757 free_list, spaceres);
2758 if (error)
2759 goto out_trans_abort;
2760
2761 /* transfer ip2 ".." reference to dp1 */
2762 if (!S_ISDIR(VFS_I(ip1)->i_mode)) {
2763 error = xfs_droplink(tp, dp2);
2764 if (error)
2765 goto out_trans_abort;
2766 error = xfs_bumplink(tp, dp1);
2767 if (error)
2768 goto out_trans_abort;
2769 }
2770
2771 /*
2772 * Although ip1 isn't changed here, userspace needs
2773 * to be warned about the change, so that applications
2774 * relying on it (like backup ones), will properly
2775 * notify the change
2776 */
2777 ip1_flags |= XFS_ICHGTIME_CHG;
2778 ip2_flags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2779 }
2780
2781 if (S_ISDIR(VFS_I(ip1)->i_mode)) {
2782 error = xfs_dir_replace(tp, ip1, &xfs_name_dotdot,
2783 dp2->i_ino, first_block,
2784 free_list, spaceres);
2785 if (error)
2786 goto out_trans_abort;
2787
2788 /* transfer ip1 ".." reference to dp2 */
2789 if (!S_ISDIR(VFS_I(ip2)->i_mode)) {
2790 error = xfs_droplink(tp, dp1);
2791 if (error)
2792 goto out_trans_abort;
2793 error = xfs_bumplink(tp, dp2);
2794 if (error)
2795 goto out_trans_abort;
2796 }
2797
2798 /*
2799 * Although ip2 isn't changed here, userspace needs
2800 * to be warned about the change, so that applications
2801 * relying on it (like backup ones), will properly
2802 * notify the change
2803 */
2804 ip1_flags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2805 ip2_flags |= XFS_ICHGTIME_CHG;
2806 }
2807 }
2808
2809 if (ip1_flags) {
2810 xfs_trans_ichgtime(tp, ip1, ip1_flags);
2811 xfs_trans_log_inode(tp, ip1, XFS_ILOG_CORE);
2812 }
2813 if (ip2_flags) {
2814 xfs_trans_ichgtime(tp, ip2, ip2_flags);
2815 xfs_trans_log_inode(tp, ip2, XFS_ILOG_CORE);
2816 }
2817 if (dp2_flags) {
2818 xfs_trans_ichgtime(tp, dp2, dp2_flags);
2819 xfs_trans_log_inode(tp, dp2, XFS_ILOG_CORE);
2820 }
2821 xfs_trans_ichgtime(tp, dp1, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2822 xfs_trans_log_inode(tp, dp1, XFS_ILOG_CORE);
2823 return xfs_finish_rename(tp, free_list);
2824
2825out_trans_abort:
2826 xfs_bmap_cancel(free_list);
2827 xfs_trans_cancel(tp);
2828 return error;
2829}
2830
2831/*
2832 * xfs_rename_alloc_whiteout()
2833 *
2834 * Return a referenced, unlinked, unlocked inode that that can be used as a
2835 * whiteout in a rename transaction. We use a tmpfile inode here so that if we
2836 * crash between allocating the inode and linking it into the rename transaction
2837 * recovery will free the inode and we won't leak it.
2838 */
2839static int
2840xfs_rename_alloc_whiteout(
2841 struct xfs_inode *dp,
2842 struct xfs_inode **wip)
2843{
2844 struct xfs_inode *tmpfile;
2845 int error;
2846
2847 error = xfs_create_tmpfile(dp, NULL, S_IFCHR | WHITEOUT_MODE, &tmpfile);
2848 if (error)
2849 return error;
2850
2851 /*
2852 * Prepare the tmpfile inode as if it were created through the VFS.
2853 * Otherwise, the link increment paths will complain about nlink 0->1.
2854 * Drop the link count as done by d_tmpfile(), complete the inode setup
2855 * and flag it as linkable.
2856 */
2857 drop_nlink(VFS_I(tmpfile));
2858 xfs_finish_inode_setup(tmpfile);
2859 VFS_I(tmpfile)->i_state |= I_LINKABLE;
2860
2861 *wip = tmpfile;
2862 return 0;
2863}
2864
2865/*
2866 * xfs_rename
2867 */
2868int
2869xfs_rename(
2870 struct xfs_inode *src_dp,
2871 struct xfs_name *src_name,
2872 struct xfs_inode *src_ip,
2873 struct xfs_inode *target_dp,
2874 struct xfs_name *target_name,
2875 struct xfs_inode *target_ip,
2876 unsigned int flags)
2877{
2878 struct xfs_mount *mp = src_dp->i_mount;
2879 struct xfs_trans *tp;
2880 struct xfs_bmap_free free_list;
2881 xfs_fsblock_t first_block;
2882 struct xfs_inode *wip = NULL; /* whiteout inode */
2883 struct xfs_inode *inodes[__XFS_SORT_INODES];
2884 int num_inodes = __XFS_SORT_INODES;
2885 bool new_parent = (src_dp != target_dp);
2886 bool src_is_directory = S_ISDIR(VFS_I(src_ip)->i_mode);
2887 int spaceres;
2888 int error;
2889
2890 trace_xfs_rename(src_dp, target_dp, src_name, target_name);
2891
2892 if ((flags & RENAME_EXCHANGE) && !target_ip)
2893 return -EINVAL;
2894
2895 /*
2896 * If we are doing a whiteout operation, allocate the whiteout inode
2897 * we will be placing at the target and ensure the type is set
2898 * appropriately.
2899 */
2900 if (flags & RENAME_WHITEOUT) {
2901 ASSERT(!(flags & (RENAME_NOREPLACE | RENAME_EXCHANGE)));
2902 error = xfs_rename_alloc_whiteout(target_dp, &wip);
2903 if (error)
2904 return error;
2905
2906 /* setup target dirent info as whiteout */
2907 src_name->type = XFS_DIR3_FT_CHRDEV;
2908 }
2909
2910 xfs_sort_for_rename(src_dp, target_dp, src_ip, target_ip, wip,
2911 inodes, &num_inodes);
2912
2913 tp = xfs_trans_alloc(mp, XFS_TRANS_RENAME);
2914 spaceres = XFS_RENAME_SPACE_RES(mp, target_name->len);
2915 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_rename, spaceres, 0);
2916 if (error == -ENOSPC) {
2917 spaceres = 0;
2918 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_rename, 0, 0);
2919 }
2920 if (error)
2921 goto out_trans_cancel;
2922
2923 /*
2924 * Attach the dquots to the inodes
2925 */
2926 error = xfs_qm_vop_rename_dqattach(inodes);
2927 if (error)
2928 goto out_trans_cancel;
2929
2930 /*
2931 * Lock all the participating inodes. Depending upon whether
2932 * the target_name exists in the target directory, and
2933 * whether the target directory is the same as the source
2934 * directory, we can lock from 2 to 4 inodes.
2935 */
2936 if (!new_parent)
2937 xfs_ilock(src_dp, XFS_IOLOCK_EXCL | XFS_IOLOCK_PARENT);
2938 else
2939 xfs_lock_two_inodes(src_dp, target_dp,
2940 XFS_IOLOCK_EXCL | XFS_IOLOCK_PARENT);
2941
2942 xfs_lock_inodes(inodes, num_inodes, XFS_ILOCK_EXCL);
2943
2944 /*
2945 * Join all the inodes to the transaction. From this point on,
2946 * we can rely on either trans_commit or trans_cancel to unlock
2947 * them.
2948 */
2949 xfs_trans_ijoin(tp, src_dp, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
2950 if (new_parent)
2951 xfs_trans_ijoin(tp, target_dp, XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL);
2952 xfs_trans_ijoin(tp, src_ip, XFS_ILOCK_EXCL);
2953 if (target_ip)
2954 xfs_trans_ijoin(tp, target_ip, XFS_ILOCK_EXCL);
2955 if (wip)
2956 xfs_trans_ijoin(tp, wip, XFS_ILOCK_EXCL);
2957
2958 /*
2959 * If we are using project inheritance, we only allow renames
2960 * into our tree when the project IDs are the same; else the
2961 * tree quota mechanism would be circumvented.
2962 */
2963 if (unlikely((target_dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
2964 (xfs_get_projid(target_dp) != xfs_get_projid(src_ip)))) {
2965 error = -EXDEV;
2966 goto out_trans_cancel;
2967 }
2968
2969 xfs_bmap_init(&free_list, &first_block);
2970
2971 /* RENAME_EXCHANGE is unique from here on. */
2972 if (flags & RENAME_EXCHANGE)
2973 return xfs_cross_rename(tp, src_dp, src_name, src_ip,
2974 target_dp, target_name, target_ip,
2975 &free_list, &first_block, spaceres);
2976
2977 /*
2978 * Set up the target.
2979 */
2980 if (target_ip == NULL) {
2981 /*
2982 * If there's no space reservation, check the entry will
2983 * fit before actually inserting it.
2984 */
2985 if (!spaceres) {
2986 error = xfs_dir_canenter(tp, target_dp, target_name);
2987 if (error)
2988 goto out_trans_cancel;
2989 }
2990 /*
2991 * If target does not exist and the rename crosses
2992 * directories, adjust the target directory link count
2993 * to account for the ".." reference from the new entry.
2994 */
2995 error = xfs_dir_createname(tp, target_dp, target_name,
2996 src_ip->i_ino, &first_block,
2997 &free_list, spaceres);
2998 if (error)
2999 goto out_bmap_cancel;
3000
3001 xfs_trans_ichgtime(tp, target_dp,
3002 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3003
3004 if (new_parent && src_is_directory) {
3005 error = xfs_bumplink(tp, target_dp);
3006 if (error)
3007 goto out_bmap_cancel;
3008 }
3009 } else { /* target_ip != NULL */
3010 /*
3011 * If target exists and it's a directory, check that both
3012 * target and source are directories and that target can be
3013 * destroyed, or that neither is a directory.
3014 */
3015 if (S_ISDIR(VFS_I(target_ip)->i_mode)) {
3016 /*
3017 * Make sure target dir is empty.
3018 */
3019 if (!(xfs_dir_isempty(target_ip)) ||
3020 (VFS_I(target_ip)->i_nlink > 2)) {
3021 error = -EEXIST;
3022 goto out_trans_cancel;
3023 }
3024 }
3025
3026 /*
3027 * Link the source inode under the target name.
3028 * If the source inode is a directory and we are moving
3029 * it across directories, its ".." entry will be
3030 * inconsistent until we replace that down below.
3031 *
3032 * In case there is already an entry with the same
3033 * name at the destination directory, remove it first.
3034 */
3035 error = xfs_dir_replace(tp, target_dp, target_name,
3036 src_ip->i_ino,
3037 &first_block, &free_list, spaceres);
3038 if (error)
3039 goto out_bmap_cancel;
3040
3041 xfs_trans_ichgtime(tp, target_dp,
3042 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3043
3044 /*
3045 * Decrement the link count on the target since the target
3046 * dir no longer points to it.
3047 */
3048 error = xfs_droplink(tp, target_ip);
3049 if (error)
3050 goto out_bmap_cancel;
3051
3052 if (src_is_directory) {
3053 /*
3054 * Drop the link from the old "." entry.
3055 */
3056 error = xfs_droplink(tp, target_ip);
3057 if (error)
3058 goto out_bmap_cancel;
3059 }
3060 } /* target_ip != NULL */
3061
3062 /*
3063 * Remove the source.
3064 */
3065 if (new_parent && src_is_directory) {
3066 /*
3067 * Rewrite the ".." entry to point to the new
3068 * directory.
3069 */
3070 error = xfs_dir_replace(tp, src_ip, &xfs_name_dotdot,
3071 target_dp->i_ino,
3072 &first_block, &free_list, spaceres);
3073 ASSERT(error != -EEXIST);
3074 if (error)
3075 goto out_bmap_cancel;
3076 }
3077
3078 /*
3079 * We always want to hit the ctime on the source inode.
3080 *
3081 * This isn't strictly required by the standards since the source
3082 * inode isn't really being changed, but old unix file systems did
3083 * it and some incremental backup programs won't work without it.
3084 */
3085 xfs_trans_ichgtime(tp, src_ip, XFS_ICHGTIME_CHG);
3086 xfs_trans_log_inode(tp, src_ip, XFS_ILOG_CORE);
3087
3088 /*
3089 * Adjust the link count on src_dp. This is necessary when
3090 * renaming a directory, either within one parent when
3091 * the target existed, or across two parent directories.
3092 */
3093 if (src_is_directory && (new_parent || target_ip != NULL)) {
3094
3095 /*
3096 * Decrement link count on src_directory since the
3097 * entry that's moved no longer points to it.
3098 */
3099 error = xfs_droplink(tp, src_dp);
3100 if (error)
3101 goto out_bmap_cancel;
3102 }
3103
3104 /*
3105 * For whiteouts, we only need to update the source dirent with the
3106 * inode number of the whiteout inode rather than removing it
3107 * altogether.
3108 */
3109 if (wip) {
3110 error = xfs_dir_replace(tp, src_dp, src_name, wip->i_ino,
3111 &first_block, &free_list, spaceres);
3112 } else
3113 error = xfs_dir_removename(tp, src_dp, src_name, src_ip->i_ino,
3114 &first_block, &free_list, spaceres);
3115 if (error)
3116 goto out_bmap_cancel;
3117
3118 /*
3119 * For whiteouts, we need to bump the link count on the whiteout inode.
3120 * This means that failures all the way up to this point leave the inode
3121 * on the unlinked list and so cleanup is a simple matter of dropping
3122 * the remaining reference to it. If we fail here after bumping the link
3123 * count, we're shutting down the filesystem so we'll never see the
3124 * intermediate state on disk.
3125 */
3126 if (wip) {
3127 ASSERT(VFS_I(wip)->i_nlink == 0);
3128 error = xfs_bumplink(tp, wip);
3129 if (error)
3130 goto out_bmap_cancel;
3131 error = xfs_iunlink_remove(tp, wip);
3132 if (error)
3133 goto out_bmap_cancel;
3134 xfs_trans_log_inode(tp, wip, XFS_ILOG_CORE);
3135
3136 /*
3137 * Now we have a real link, clear the "I'm a tmpfile" state
3138 * flag from the inode so it doesn't accidentally get misused in
3139 * future.
3140 */
3141 VFS_I(wip)->i_state &= ~I_LINKABLE;
3142 }
3143
3144 xfs_trans_ichgtime(tp, src_dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3145 xfs_trans_log_inode(tp, src_dp, XFS_ILOG_CORE);
3146 if (new_parent)
3147 xfs_trans_log_inode(tp, target_dp, XFS_ILOG_CORE);
3148
3149 error = xfs_finish_rename(tp, &free_list);
3150 if (wip)
3151 IRELE(wip);
3152 return error;
3153
3154out_bmap_cancel:
3155 xfs_bmap_cancel(&free_list);
3156out_trans_cancel:
3157 xfs_trans_cancel(tp);
3158 if (wip)
3159 IRELE(wip);
3160 return error;
3161}
3162
3163STATIC int
3164xfs_iflush_cluster(
3165 xfs_inode_t *ip,
3166 xfs_buf_t *bp)
3167{
3168 xfs_mount_t *mp = ip->i_mount;
3169 struct xfs_perag *pag;
3170 unsigned long first_index, mask;
3171 unsigned long inodes_per_cluster;
3172 int ilist_size;
3173 xfs_inode_t **ilist;
3174 xfs_inode_t *iq;
3175 int nr_found;
3176 int clcount = 0;
3177 int bufwasdelwri;
3178 int i;
3179
3180 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
3181
3182 inodes_per_cluster = mp->m_inode_cluster_size >> mp->m_sb.sb_inodelog;
3183 ilist_size = inodes_per_cluster * sizeof(xfs_inode_t *);
3184 ilist = kmem_alloc(ilist_size, KM_MAYFAIL|KM_NOFS);
3185 if (!ilist)
3186 goto out_put;
3187
3188 mask = ~(((mp->m_inode_cluster_size >> mp->m_sb.sb_inodelog)) - 1);
3189 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino) & mask;
3190 rcu_read_lock();
3191 /* really need a gang lookup range call here */
3192 nr_found = radix_tree_gang_lookup(&pag->pag_ici_root, (void**)ilist,
3193 first_index, inodes_per_cluster);
3194 if (nr_found == 0)
3195 goto out_free;
3196
3197 for (i = 0; i < nr_found; i++) {
3198 iq = ilist[i];
3199 if (iq == ip)
3200 continue;
3201
3202 /*
3203 * because this is an RCU protected lookup, we could find a
3204 * recently freed or even reallocated inode during the lookup.
3205 * We need to check under the i_flags_lock for a valid inode
3206 * here. Skip it if it is not valid or the wrong inode.
3207 */
3208 spin_lock(&ip->i_flags_lock);
3209 if (!ip->i_ino ||
3210 (XFS_INO_TO_AGINO(mp, iq->i_ino) & mask) != first_index) {
3211 spin_unlock(&ip->i_flags_lock);
3212 continue;
3213 }
3214 spin_unlock(&ip->i_flags_lock);
3215
3216 /*
3217 * Do an un-protected check to see if the inode is dirty and
3218 * is a candidate for flushing. These checks will be repeated
3219 * later after the appropriate locks are acquired.
3220 */
3221 if (xfs_inode_clean(iq) && xfs_ipincount(iq) == 0)
3222 continue;
3223
3224 /*
3225 * Try to get locks. If any are unavailable or it is pinned,
3226 * then this inode cannot be flushed and is skipped.
3227 */
3228
3229 if (!xfs_ilock_nowait(iq, XFS_ILOCK_SHARED))
3230 continue;
3231 if (!xfs_iflock_nowait(iq)) {
3232 xfs_iunlock(iq, XFS_ILOCK_SHARED);
3233 continue;
3234 }
3235 if (xfs_ipincount(iq)) {
3236 xfs_ifunlock(iq);
3237 xfs_iunlock(iq, XFS_ILOCK_SHARED);
3238 continue;
3239 }
3240
3241 /*
3242 * arriving here means that this inode can be flushed. First
3243 * re-check that it's dirty before flushing.
3244 */
3245 if (!xfs_inode_clean(iq)) {
3246 int error;
3247 error = xfs_iflush_int(iq, bp);
3248 if (error) {
3249 xfs_iunlock(iq, XFS_ILOCK_SHARED);
3250 goto cluster_corrupt_out;
3251 }
3252 clcount++;
3253 } else {
3254 xfs_ifunlock(iq);
3255 }
3256 xfs_iunlock(iq, XFS_ILOCK_SHARED);
3257 }
3258
3259 if (clcount) {
3260 XFS_STATS_INC(mp, xs_icluster_flushcnt);
3261 XFS_STATS_ADD(mp, xs_icluster_flushinode, clcount);
3262 }
3263
3264out_free:
3265 rcu_read_unlock();
3266 kmem_free(ilist);
3267out_put:
3268 xfs_perag_put(pag);
3269 return 0;
3270
3271
3272cluster_corrupt_out:
3273 /*
3274 * Corruption detected in the clustering loop. Invalidate the
3275 * inode buffer and shut down the filesystem.
3276 */
3277 rcu_read_unlock();
3278 /*
3279 * Clean up the buffer. If it was delwri, just release it --
3280 * brelse can handle it with no problems. If not, shut down the
3281 * filesystem before releasing the buffer.
3282 */
3283 bufwasdelwri = (bp->b_flags & _XBF_DELWRI_Q);
3284 if (bufwasdelwri)
3285 xfs_buf_relse(bp);
3286
3287 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
3288
3289 if (!bufwasdelwri) {
3290 /*
3291 * Just like incore_relse: if we have b_iodone functions,
3292 * mark the buffer as an error and call them. Otherwise
3293 * mark it as stale and brelse.
3294 */
3295 if (bp->b_iodone) {
3296 bp->b_flags &= ~XBF_DONE;
3297 xfs_buf_stale(bp);
3298 xfs_buf_ioerror(bp, -EIO);
3299 xfs_buf_ioend(bp);
3300 } else {
3301 xfs_buf_stale(bp);
3302 xfs_buf_relse(bp);
3303 }
3304 }
3305
3306 /*
3307 * Unlocks the flush lock
3308 */
3309 xfs_iflush_abort(iq, false);
3310 kmem_free(ilist);
3311 xfs_perag_put(pag);
3312 return -EFSCORRUPTED;
3313}
3314
3315/*
3316 * Flush dirty inode metadata into the backing buffer.
3317 *
3318 * The caller must have the inode lock and the inode flush lock held. The
3319 * inode lock will still be held upon return to the caller, and the inode
3320 * flush lock will be released after the inode has reached the disk.
3321 *
3322 * The caller must write out the buffer returned in *bpp and release it.
3323 */
3324int
3325xfs_iflush(
3326 struct xfs_inode *ip,
3327 struct xfs_buf **bpp)
3328{
3329 struct xfs_mount *mp = ip->i_mount;
3330 struct xfs_buf *bp;
3331 struct xfs_dinode *dip;
3332 int error;
3333
3334 XFS_STATS_INC(mp, xs_iflush_count);
3335
3336 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
3337 ASSERT(xfs_isiflocked(ip));
3338 ASSERT(ip->i_d.di_format != XFS_DINODE_FMT_BTREE ||
3339 ip->i_d.di_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
3340
3341 *bpp = NULL;
3342
3343 xfs_iunpin_wait(ip);
3344
3345 /*
3346 * For stale inodes we cannot rely on the backing buffer remaining
3347 * stale in cache for the remaining life of the stale inode and so
3348 * xfs_imap_to_bp() below may give us a buffer that no longer contains
3349 * inodes below. We have to check this after ensuring the inode is
3350 * unpinned so that it is safe to reclaim the stale inode after the
3351 * flush call.
3352 */
3353 if (xfs_iflags_test(ip, XFS_ISTALE)) {
3354 xfs_ifunlock(ip);
3355 return 0;
3356 }
3357
3358 /*
3359 * This may have been unpinned because the filesystem is shutting
3360 * down forcibly. If that's the case we must not write this inode
3361 * to disk, because the log record didn't make it to disk.
3362 *
3363 * We also have to remove the log item from the AIL in this case,
3364 * as we wait for an empty AIL as part of the unmount process.
3365 */
3366 if (XFS_FORCED_SHUTDOWN(mp)) {
3367 error = -EIO;
3368 goto abort_out;
3369 }
3370
3371 /*
3372 * Get the buffer containing the on-disk inode.
3373 */
3374 error = xfs_imap_to_bp(mp, NULL, &ip->i_imap, &dip, &bp, XBF_TRYLOCK,
3375 0);
3376 if (error || !bp) {
3377 xfs_ifunlock(ip);
3378 return error;
3379 }
3380
3381 /*
3382 * First flush out the inode that xfs_iflush was called with.
3383 */
3384 error = xfs_iflush_int(ip, bp);
3385 if (error)
3386 goto corrupt_out;
3387
3388 /*
3389 * If the buffer is pinned then push on the log now so we won't
3390 * get stuck waiting in the write for too long.
3391 */
3392 if (xfs_buf_ispinned(bp))
3393 xfs_log_force(mp, 0);
3394
3395 /*
3396 * inode clustering:
3397 * see if other inodes can be gathered into this write
3398 */
3399 error = xfs_iflush_cluster(ip, bp);
3400 if (error)
3401 goto cluster_corrupt_out;
3402
3403 *bpp = bp;
3404 return 0;
3405
3406corrupt_out:
3407 xfs_buf_relse(bp);
3408 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
3409cluster_corrupt_out:
3410 error = -EFSCORRUPTED;
3411abort_out:
3412 /*
3413 * Unlocks the flush lock
3414 */
3415 xfs_iflush_abort(ip, false);
3416 return error;
3417}
3418
3419STATIC int
3420xfs_iflush_int(
3421 struct xfs_inode *ip,
3422 struct xfs_buf *bp)
3423{
3424 struct xfs_inode_log_item *iip = ip->i_itemp;
3425 struct xfs_dinode *dip;
3426 struct xfs_mount *mp = ip->i_mount;
3427
3428 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
3429 ASSERT(xfs_isiflocked(ip));
3430 ASSERT(ip->i_d.di_format != XFS_DINODE_FMT_BTREE ||
3431 ip->i_d.di_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
3432 ASSERT(iip != NULL && iip->ili_fields != 0);
3433 ASSERT(ip->i_d.di_version > 1);
3434
3435 /* set *dip = inode's place in the buffer */
3436 dip = xfs_buf_offset(bp, ip->i_imap.im_boffset);
3437
3438 if (XFS_TEST_ERROR(dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC),
3439 mp, XFS_ERRTAG_IFLUSH_1, XFS_RANDOM_IFLUSH_1)) {
3440 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3441 "%s: Bad inode %Lu magic number 0x%x, ptr 0x%p",
3442 __func__, ip->i_ino, be16_to_cpu(dip->di_magic), dip);
3443 goto corrupt_out;
3444 }
3445 if (S_ISREG(VFS_I(ip)->i_mode)) {
3446 if (XFS_TEST_ERROR(
3447 (ip->i_d.di_format != XFS_DINODE_FMT_EXTENTS) &&
3448 (ip->i_d.di_format != XFS_DINODE_FMT_BTREE),
3449 mp, XFS_ERRTAG_IFLUSH_3, XFS_RANDOM_IFLUSH_3)) {
3450 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3451 "%s: Bad regular inode %Lu, ptr 0x%p",
3452 __func__, ip->i_ino, ip);
3453 goto corrupt_out;
3454 }
3455 } else if (S_ISDIR(VFS_I(ip)->i_mode)) {
3456 if (XFS_TEST_ERROR(
3457 (ip->i_d.di_format != XFS_DINODE_FMT_EXTENTS) &&
3458 (ip->i_d.di_format != XFS_DINODE_FMT_BTREE) &&
3459 (ip->i_d.di_format != XFS_DINODE_FMT_LOCAL),
3460 mp, XFS_ERRTAG_IFLUSH_4, XFS_RANDOM_IFLUSH_4)) {
3461 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3462 "%s: Bad directory inode %Lu, ptr 0x%p",
3463 __func__, ip->i_ino, ip);
3464 goto corrupt_out;
3465 }
3466 }
3467 if (XFS_TEST_ERROR(ip->i_d.di_nextents + ip->i_d.di_anextents >
3468 ip->i_d.di_nblocks, mp, XFS_ERRTAG_IFLUSH_5,
3469 XFS_RANDOM_IFLUSH_5)) {
3470 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3471 "%s: detected corrupt incore inode %Lu, "
3472 "total extents = %d, nblocks = %Ld, ptr 0x%p",
3473 __func__, ip->i_ino,
3474 ip->i_d.di_nextents + ip->i_d.di_anextents,
3475 ip->i_d.di_nblocks, ip);
3476 goto corrupt_out;
3477 }
3478 if (XFS_TEST_ERROR(ip->i_d.di_forkoff > mp->m_sb.sb_inodesize,
3479 mp, XFS_ERRTAG_IFLUSH_6, XFS_RANDOM_IFLUSH_6)) {
3480 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3481 "%s: bad inode %Lu, forkoff 0x%x, ptr 0x%p",
3482 __func__, ip->i_ino, ip->i_d.di_forkoff, ip);
3483 goto corrupt_out;
3484 }
3485
3486 /*
3487 * Inode item log recovery for v2 inodes are dependent on the
3488 * di_flushiter count for correct sequencing. We bump the flush
3489 * iteration count so we can detect flushes which postdate a log record
3490 * during recovery. This is redundant as we now log every change and
3491 * hence this can't happen but we need to still do it to ensure
3492 * backwards compatibility with old kernels that predate logging all
3493 * inode changes.
3494 */
3495 if (ip->i_d.di_version < 3)
3496 ip->i_d.di_flushiter++;
3497
3498 /*
3499 * Copy the dirty parts of the inode into the on-disk inode. We always
3500 * copy out the core of the inode, because if the inode is dirty at all
3501 * the core must be.
3502 */
3503 xfs_inode_to_disk(ip, dip, iip->ili_item.li_lsn);
3504
3505 /* Wrap, we never let the log put out DI_MAX_FLUSH */
3506 if (ip->i_d.di_flushiter == DI_MAX_FLUSH)
3507 ip->i_d.di_flushiter = 0;
3508
3509 xfs_iflush_fork(ip, dip, iip, XFS_DATA_FORK);
3510 if (XFS_IFORK_Q(ip))
3511 xfs_iflush_fork(ip, dip, iip, XFS_ATTR_FORK);
3512 xfs_inobp_check(mp, bp);
3513
3514 /*
3515 * We've recorded everything logged in the inode, so we'd like to clear
3516 * the ili_fields bits so we don't log and flush things unnecessarily.
3517 * However, we can't stop logging all this information until the data
3518 * we've copied into the disk buffer is written to disk. If we did we
3519 * might overwrite the copy of the inode in the log with all the data
3520 * after re-logging only part of it, and in the face of a crash we
3521 * wouldn't have all the data we need to recover.
3522 *
3523 * What we do is move the bits to the ili_last_fields field. When
3524 * logging the inode, these bits are moved back to the ili_fields field.
3525 * In the xfs_iflush_done() routine we clear ili_last_fields, since we
3526 * know that the information those bits represent is permanently on
3527 * disk. As long as the flush completes before the inode is logged
3528 * again, then both ili_fields and ili_last_fields will be cleared.
3529 *
3530 * We can play with the ili_fields bits here, because the inode lock
3531 * must be held exclusively in order to set bits there and the flush
3532 * lock protects the ili_last_fields bits. Set ili_logged so the flush
3533 * done routine can tell whether or not to look in the AIL. Also, store
3534 * the current LSN of the inode so that we can tell whether the item has
3535 * moved in the AIL from xfs_iflush_done(). In order to read the lsn we
3536 * need the AIL lock, because it is a 64 bit value that cannot be read
3537 * atomically.
3538 */
3539 iip->ili_last_fields = iip->ili_fields;
3540 iip->ili_fields = 0;
3541 iip->ili_fsync_fields = 0;
3542 iip->ili_logged = 1;
3543
3544 xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
3545 &iip->ili_item.li_lsn);
3546
3547 /*
3548 * Attach the function xfs_iflush_done to the inode's
3549 * buffer. This will remove the inode from the AIL
3550 * and unlock the inode's flush lock when the inode is
3551 * completely written to disk.
3552 */
3553 xfs_buf_attach_iodone(bp, xfs_iflush_done, &iip->ili_item);
3554
3555 /* generate the checksum. */
3556 xfs_dinode_calc_crc(mp, dip);
3557
3558 ASSERT(bp->b_fspriv != NULL);
3559 ASSERT(bp->b_iodone != NULL);
3560 return 0;
3561
3562corrupt_out:
3563 return -EFSCORRUPTED;
3564}