Loading...
1/*
2 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
3 * All Rights Reserved.
4 *
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
8 *
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
13 *
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
17 */
18#include <linux/log2.h>
19
20#include "xfs.h"
21#include "xfs_fs.h"
22#include "xfs_shared.h"
23#include "xfs_format.h"
24#include "xfs_log_format.h"
25#include "xfs_trans_resv.h"
26#include "xfs_sb.h"
27#include "xfs_mount.h"
28#include "xfs_defer.h"
29#include "xfs_inode.h"
30#include "xfs_da_format.h"
31#include "xfs_da_btree.h"
32#include "xfs_dir2.h"
33#include "xfs_attr_sf.h"
34#include "xfs_attr.h"
35#include "xfs_trans_space.h"
36#include "xfs_trans.h"
37#include "xfs_buf_item.h"
38#include "xfs_inode_item.h"
39#include "xfs_ialloc.h"
40#include "xfs_bmap.h"
41#include "xfs_bmap_util.h"
42#include "xfs_error.h"
43#include "xfs_quota.h"
44#include "xfs_filestream.h"
45#include "xfs_cksum.h"
46#include "xfs_trace.h"
47#include "xfs_icache.h"
48#include "xfs_symlink.h"
49#include "xfs_trans_priv.h"
50#include "xfs_log.h"
51#include "xfs_bmap_btree.h"
52#include "xfs_reflink.h"
53
54kmem_zone_t *xfs_inode_zone;
55
56/*
57 * Used in xfs_itruncate_extents(). This is the maximum number of extents
58 * freed from a file in a single transaction.
59 */
60#define XFS_ITRUNC_MAX_EXTENTS 2
61
62STATIC int xfs_iflush_int(struct xfs_inode *, struct xfs_buf *);
63STATIC int xfs_iunlink(struct xfs_trans *, struct xfs_inode *);
64STATIC int xfs_iunlink_remove(struct xfs_trans *, struct xfs_inode *);
65
66/*
67 * helper function to extract extent size hint from inode
68 */
69xfs_extlen_t
70xfs_get_extsz_hint(
71 struct xfs_inode *ip)
72{
73 if ((ip->i_d.di_flags & XFS_DIFLAG_EXTSIZE) && ip->i_d.di_extsize)
74 return ip->i_d.di_extsize;
75 if (XFS_IS_REALTIME_INODE(ip))
76 return ip->i_mount->m_sb.sb_rextsize;
77 return 0;
78}
79
80/*
81 * Helper function to extract CoW extent size hint from inode.
82 * Between the extent size hint and the CoW extent size hint, we
83 * return the greater of the two. If the value is zero (automatic),
84 * use the default size.
85 */
86xfs_extlen_t
87xfs_get_cowextsz_hint(
88 struct xfs_inode *ip)
89{
90 xfs_extlen_t a, b;
91
92 a = 0;
93 if (ip->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE)
94 a = ip->i_d.di_cowextsize;
95 b = xfs_get_extsz_hint(ip);
96
97 a = max(a, b);
98 if (a == 0)
99 return XFS_DEFAULT_COWEXTSZ_HINT;
100 return a;
101}
102
103/*
104 * These two are wrapper routines around the xfs_ilock() routine used to
105 * centralize some grungy code. They are used in places that wish to lock the
106 * inode solely for reading the extents. The reason these places can't just
107 * call xfs_ilock(ip, XFS_ILOCK_SHARED) is that the inode lock also guards to
108 * bringing in of the extents from disk for a file in b-tree format. If the
109 * inode is in b-tree format, then we need to lock the inode exclusively until
110 * the extents are read in. Locking it exclusively all the time would limit
111 * our parallelism unnecessarily, though. What we do instead is check to see
112 * if the extents have been read in yet, and only lock the inode exclusively
113 * if they have not.
114 *
115 * The functions return a value which should be given to the corresponding
116 * xfs_iunlock() call.
117 */
118uint
119xfs_ilock_data_map_shared(
120 struct xfs_inode *ip)
121{
122 uint lock_mode = XFS_ILOCK_SHARED;
123
124 if (ip->i_d.di_format == XFS_DINODE_FMT_BTREE &&
125 (ip->i_df.if_flags & XFS_IFEXTENTS) == 0)
126 lock_mode = XFS_ILOCK_EXCL;
127 xfs_ilock(ip, lock_mode);
128 return lock_mode;
129}
130
131uint
132xfs_ilock_attr_map_shared(
133 struct xfs_inode *ip)
134{
135 uint lock_mode = XFS_ILOCK_SHARED;
136
137 if (ip->i_d.di_aformat == XFS_DINODE_FMT_BTREE &&
138 (ip->i_afp->if_flags & XFS_IFEXTENTS) == 0)
139 lock_mode = XFS_ILOCK_EXCL;
140 xfs_ilock(ip, lock_mode);
141 return lock_mode;
142}
143
144/*
145 * In addition to i_rwsem in the VFS inode, the xfs inode contains 2
146 * multi-reader locks: i_mmap_lock and the i_lock. This routine allows
147 * various combinations of the locks to be obtained.
148 *
149 * The 3 locks should always be ordered so that the IO lock is obtained first,
150 * the mmap lock second and the ilock last in order to prevent deadlock.
151 *
152 * Basic locking order:
153 *
154 * i_rwsem -> i_mmap_lock -> page_lock -> i_ilock
155 *
156 * mmap_sem locking order:
157 *
158 * i_rwsem -> page lock -> mmap_sem
159 * mmap_sem -> i_mmap_lock -> page_lock
160 *
161 * The difference in mmap_sem locking order mean that we cannot hold the
162 * i_mmap_lock over syscall based read(2)/write(2) based IO. These IO paths can
163 * fault in pages during copy in/out (for buffered IO) or require the mmap_sem
164 * in get_user_pages() to map the user pages into the kernel address space for
165 * direct IO. Similarly the i_rwsem cannot be taken inside a page fault because
166 * page faults already hold the mmap_sem.
167 *
168 * Hence to serialise fully against both syscall and mmap based IO, we need to
169 * take both the i_rwsem and the i_mmap_lock. These locks should *only* be both
170 * taken in places where we need to invalidate the page cache in a race
171 * free manner (e.g. truncate, hole punch and other extent manipulation
172 * functions).
173 */
174void
175xfs_ilock(
176 xfs_inode_t *ip,
177 uint lock_flags)
178{
179 trace_xfs_ilock(ip, lock_flags, _RET_IP_);
180
181 /*
182 * You can't set both SHARED and EXCL for the same lock,
183 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
184 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
185 */
186 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
187 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
188 ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
189 (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
190 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
191 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
192 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
193
194 if (lock_flags & XFS_IOLOCK_EXCL) {
195 down_write_nested(&VFS_I(ip)->i_rwsem,
196 XFS_IOLOCK_DEP(lock_flags));
197 } else if (lock_flags & XFS_IOLOCK_SHARED) {
198 down_read_nested(&VFS_I(ip)->i_rwsem,
199 XFS_IOLOCK_DEP(lock_flags));
200 }
201
202 if (lock_flags & XFS_MMAPLOCK_EXCL)
203 mrupdate_nested(&ip->i_mmaplock, XFS_MMAPLOCK_DEP(lock_flags));
204 else if (lock_flags & XFS_MMAPLOCK_SHARED)
205 mraccess_nested(&ip->i_mmaplock, XFS_MMAPLOCK_DEP(lock_flags));
206
207 if (lock_flags & XFS_ILOCK_EXCL)
208 mrupdate_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
209 else if (lock_flags & XFS_ILOCK_SHARED)
210 mraccess_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
211}
212
213/*
214 * This is just like xfs_ilock(), except that the caller
215 * is guaranteed not to sleep. It returns 1 if it gets
216 * the requested locks and 0 otherwise. If the IO lock is
217 * obtained but the inode lock cannot be, then the IO lock
218 * is dropped before returning.
219 *
220 * ip -- the inode being locked
221 * lock_flags -- this parameter indicates the inode's locks to be
222 * to be locked. See the comment for xfs_ilock() for a list
223 * of valid values.
224 */
225int
226xfs_ilock_nowait(
227 xfs_inode_t *ip,
228 uint lock_flags)
229{
230 trace_xfs_ilock_nowait(ip, lock_flags, _RET_IP_);
231
232 /*
233 * You can't set both SHARED and EXCL for the same lock,
234 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
235 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
236 */
237 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
238 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
239 ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
240 (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
241 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
242 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
243 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
244
245 if (lock_flags & XFS_IOLOCK_EXCL) {
246 if (!down_write_trylock(&VFS_I(ip)->i_rwsem))
247 goto out;
248 } else if (lock_flags & XFS_IOLOCK_SHARED) {
249 if (!down_read_trylock(&VFS_I(ip)->i_rwsem))
250 goto out;
251 }
252
253 if (lock_flags & XFS_MMAPLOCK_EXCL) {
254 if (!mrtryupdate(&ip->i_mmaplock))
255 goto out_undo_iolock;
256 } else if (lock_flags & XFS_MMAPLOCK_SHARED) {
257 if (!mrtryaccess(&ip->i_mmaplock))
258 goto out_undo_iolock;
259 }
260
261 if (lock_flags & XFS_ILOCK_EXCL) {
262 if (!mrtryupdate(&ip->i_lock))
263 goto out_undo_mmaplock;
264 } else if (lock_flags & XFS_ILOCK_SHARED) {
265 if (!mrtryaccess(&ip->i_lock))
266 goto out_undo_mmaplock;
267 }
268 return 1;
269
270out_undo_mmaplock:
271 if (lock_flags & XFS_MMAPLOCK_EXCL)
272 mrunlock_excl(&ip->i_mmaplock);
273 else if (lock_flags & XFS_MMAPLOCK_SHARED)
274 mrunlock_shared(&ip->i_mmaplock);
275out_undo_iolock:
276 if (lock_flags & XFS_IOLOCK_EXCL)
277 up_write(&VFS_I(ip)->i_rwsem);
278 else if (lock_flags & XFS_IOLOCK_SHARED)
279 up_read(&VFS_I(ip)->i_rwsem);
280out:
281 return 0;
282}
283
284/*
285 * xfs_iunlock() is used to drop the inode locks acquired with
286 * xfs_ilock() and xfs_ilock_nowait(). The caller must pass
287 * in the flags given to xfs_ilock() or xfs_ilock_nowait() so
288 * that we know which locks to drop.
289 *
290 * ip -- the inode being unlocked
291 * lock_flags -- this parameter indicates the inode's locks to be
292 * to be unlocked. See the comment for xfs_ilock() for a list
293 * of valid values for this parameter.
294 *
295 */
296void
297xfs_iunlock(
298 xfs_inode_t *ip,
299 uint lock_flags)
300{
301 /*
302 * You can't set both SHARED and EXCL for the same lock,
303 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
304 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
305 */
306 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
307 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
308 ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
309 (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
310 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
311 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
312 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
313 ASSERT(lock_flags != 0);
314
315 if (lock_flags & XFS_IOLOCK_EXCL)
316 up_write(&VFS_I(ip)->i_rwsem);
317 else if (lock_flags & XFS_IOLOCK_SHARED)
318 up_read(&VFS_I(ip)->i_rwsem);
319
320 if (lock_flags & XFS_MMAPLOCK_EXCL)
321 mrunlock_excl(&ip->i_mmaplock);
322 else if (lock_flags & XFS_MMAPLOCK_SHARED)
323 mrunlock_shared(&ip->i_mmaplock);
324
325 if (lock_flags & XFS_ILOCK_EXCL)
326 mrunlock_excl(&ip->i_lock);
327 else if (lock_flags & XFS_ILOCK_SHARED)
328 mrunlock_shared(&ip->i_lock);
329
330 trace_xfs_iunlock(ip, lock_flags, _RET_IP_);
331}
332
333/*
334 * give up write locks. the i/o lock cannot be held nested
335 * if it is being demoted.
336 */
337void
338xfs_ilock_demote(
339 xfs_inode_t *ip,
340 uint lock_flags)
341{
342 ASSERT(lock_flags & (XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL));
343 ASSERT((lock_flags &
344 ~(XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL)) == 0);
345
346 if (lock_flags & XFS_ILOCK_EXCL)
347 mrdemote(&ip->i_lock);
348 if (lock_flags & XFS_MMAPLOCK_EXCL)
349 mrdemote(&ip->i_mmaplock);
350 if (lock_flags & XFS_IOLOCK_EXCL)
351 downgrade_write(&VFS_I(ip)->i_rwsem);
352
353 trace_xfs_ilock_demote(ip, lock_flags, _RET_IP_);
354}
355
356#if defined(DEBUG) || defined(XFS_WARN)
357int
358xfs_isilocked(
359 xfs_inode_t *ip,
360 uint lock_flags)
361{
362 if (lock_flags & (XFS_ILOCK_EXCL|XFS_ILOCK_SHARED)) {
363 if (!(lock_flags & XFS_ILOCK_SHARED))
364 return !!ip->i_lock.mr_writer;
365 return rwsem_is_locked(&ip->i_lock.mr_lock);
366 }
367
368 if (lock_flags & (XFS_MMAPLOCK_EXCL|XFS_MMAPLOCK_SHARED)) {
369 if (!(lock_flags & XFS_MMAPLOCK_SHARED))
370 return !!ip->i_mmaplock.mr_writer;
371 return rwsem_is_locked(&ip->i_mmaplock.mr_lock);
372 }
373
374 if (lock_flags & (XFS_IOLOCK_EXCL|XFS_IOLOCK_SHARED)) {
375 if (!(lock_flags & XFS_IOLOCK_SHARED))
376 return !debug_locks ||
377 lockdep_is_held_type(&VFS_I(ip)->i_rwsem, 0);
378 return rwsem_is_locked(&VFS_I(ip)->i_rwsem);
379 }
380
381 ASSERT(0);
382 return 0;
383}
384#endif
385
386#ifdef DEBUG
387int xfs_locked_n;
388int xfs_small_retries;
389int xfs_middle_retries;
390int xfs_lots_retries;
391int xfs_lock_delays;
392#endif
393
394/*
395 * xfs_lockdep_subclass_ok() is only used in an ASSERT, so is only called when
396 * DEBUG or XFS_WARN is set. And MAX_LOCKDEP_SUBCLASSES is then only defined
397 * when CONFIG_LOCKDEP is set. Hence the complex define below to avoid build
398 * errors and warnings.
399 */
400#if (defined(DEBUG) || defined(XFS_WARN)) && defined(CONFIG_LOCKDEP)
401static bool
402xfs_lockdep_subclass_ok(
403 int subclass)
404{
405 return subclass < MAX_LOCKDEP_SUBCLASSES;
406}
407#else
408#define xfs_lockdep_subclass_ok(subclass) (true)
409#endif
410
411/*
412 * Bump the subclass so xfs_lock_inodes() acquires each lock with a different
413 * value. This can be called for any type of inode lock combination, including
414 * parent locking. Care must be taken to ensure we don't overrun the subclass
415 * storage fields in the class mask we build.
416 */
417static inline int
418xfs_lock_inumorder(int lock_mode, int subclass)
419{
420 int class = 0;
421
422 ASSERT(!(lock_mode & (XFS_ILOCK_PARENT | XFS_ILOCK_RTBITMAP |
423 XFS_ILOCK_RTSUM)));
424 ASSERT(xfs_lockdep_subclass_ok(subclass));
425
426 if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)) {
427 ASSERT(subclass <= XFS_IOLOCK_MAX_SUBCLASS);
428 class += subclass << XFS_IOLOCK_SHIFT;
429 }
430
431 if (lock_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) {
432 ASSERT(subclass <= XFS_MMAPLOCK_MAX_SUBCLASS);
433 class += subclass << XFS_MMAPLOCK_SHIFT;
434 }
435
436 if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) {
437 ASSERT(subclass <= XFS_ILOCK_MAX_SUBCLASS);
438 class += subclass << XFS_ILOCK_SHIFT;
439 }
440
441 return (lock_mode & ~XFS_LOCK_SUBCLASS_MASK) | class;
442}
443
444/*
445 * The following routine will lock n inodes in exclusive mode. We assume the
446 * caller calls us with the inodes in i_ino order.
447 *
448 * We need to detect deadlock where an inode that we lock is in the AIL and we
449 * start waiting for another inode that is locked by a thread in a long running
450 * transaction (such as truncate). This can result in deadlock since the long
451 * running trans might need to wait for the inode we just locked in order to
452 * push the tail and free space in the log.
453 *
454 * xfs_lock_inodes() can only be used to lock one type of lock at a time -
455 * the iolock, the mmaplock or the ilock, but not more than one at a time. If we
456 * lock more than one at a time, lockdep will report false positives saying we
457 * have violated locking orders.
458 */
459static void
460xfs_lock_inodes(
461 xfs_inode_t **ips,
462 int inodes,
463 uint lock_mode)
464{
465 int attempts = 0, i, j, try_lock;
466 xfs_log_item_t *lp;
467
468 /*
469 * Currently supports between 2 and 5 inodes with exclusive locking. We
470 * support an arbitrary depth of locking here, but absolute limits on
471 * inodes depend on the the type of locking and the limits placed by
472 * lockdep annotations in xfs_lock_inumorder. These are all checked by
473 * the asserts.
474 */
475 ASSERT(ips && inodes >= 2 && inodes <= 5);
476 ASSERT(lock_mode & (XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL |
477 XFS_ILOCK_EXCL));
478 ASSERT(!(lock_mode & (XFS_IOLOCK_SHARED | XFS_MMAPLOCK_SHARED |
479 XFS_ILOCK_SHARED)));
480 ASSERT(!(lock_mode & XFS_MMAPLOCK_EXCL) ||
481 inodes <= XFS_MMAPLOCK_MAX_SUBCLASS + 1);
482 ASSERT(!(lock_mode & XFS_ILOCK_EXCL) ||
483 inodes <= XFS_ILOCK_MAX_SUBCLASS + 1);
484
485 if (lock_mode & XFS_IOLOCK_EXCL) {
486 ASSERT(!(lock_mode & (XFS_MMAPLOCK_EXCL | XFS_ILOCK_EXCL)));
487 } else if (lock_mode & XFS_MMAPLOCK_EXCL)
488 ASSERT(!(lock_mode & XFS_ILOCK_EXCL));
489
490 try_lock = 0;
491 i = 0;
492again:
493 for (; i < inodes; i++) {
494 ASSERT(ips[i]);
495
496 if (i && (ips[i] == ips[i - 1])) /* Already locked */
497 continue;
498
499 /*
500 * If try_lock is not set yet, make sure all locked inodes are
501 * not in the AIL. If any are, set try_lock to be used later.
502 */
503 if (!try_lock) {
504 for (j = (i - 1); j >= 0 && !try_lock; j--) {
505 lp = (xfs_log_item_t *)ips[j]->i_itemp;
506 if (lp && (lp->li_flags & XFS_LI_IN_AIL))
507 try_lock++;
508 }
509 }
510
511 /*
512 * If any of the previous locks we have locked is in the AIL,
513 * we must TRY to get the second and subsequent locks. If
514 * we can't get any, we must release all we have
515 * and try again.
516 */
517 if (!try_lock) {
518 xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
519 continue;
520 }
521
522 /* try_lock means we have an inode locked that is in the AIL. */
523 ASSERT(i != 0);
524 if (xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i)))
525 continue;
526
527 /*
528 * Unlock all previous guys and try again. xfs_iunlock will try
529 * to push the tail if the inode is in the AIL.
530 */
531 attempts++;
532 for (j = i - 1; j >= 0; j--) {
533 /*
534 * Check to see if we've already unlocked this one. Not
535 * the first one going back, and the inode ptr is the
536 * same.
537 */
538 if (j != (i - 1) && ips[j] == ips[j + 1])
539 continue;
540
541 xfs_iunlock(ips[j], lock_mode);
542 }
543
544 if ((attempts % 5) == 0) {
545 delay(1); /* Don't just spin the CPU */
546#ifdef DEBUG
547 xfs_lock_delays++;
548#endif
549 }
550 i = 0;
551 try_lock = 0;
552 goto again;
553 }
554
555#ifdef DEBUG
556 if (attempts) {
557 if (attempts < 5) xfs_small_retries++;
558 else if (attempts < 100) xfs_middle_retries++;
559 else xfs_lots_retries++;
560 } else {
561 xfs_locked_n++;
562 }
563#endif
564}
565
566/*
567 * xfs_lock_two_inodes() can only be used to lock one type of lock at a time -
568 * the iolock, the mmaplock or the ilock, but not more than one at a time. If we
569 * lock more than one at a time, lockdep will report false positives saying we
570 * have violated locking orders.
571 */
572void
573xfs_lock_two_inodes(
574 xfs_inode_t *ip0,
575 xfs_inode_t *ip1,
576 uint lock_mode)
577{
578 xfs_inode_t *temp;
579 int attempts = 0;
580 xfs_log_item_t *lp;
581
582 ASSERT(!(lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
583 if (lock_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL))
584 ASSERT(!(lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)));
585
586 ASSERT(ip0->i_ino != ip1->i_ino);
587
588 if (ip0->i_ino > ip1->i_ino) {
589 temp = ip0;
590 ip0 = ip1;
591 ip1 = temp;
592 }
593
594 again:
595 xfs_ilock(ip0, xfs_lock_inumorder(lock_mode, 0));
596
597 /*
598 * If the first lock we have locked is in the AIL, we must TRY to get
599 * the second lock. If we can't get it, we must release the first one
600 * and try again.
601 */
602 lp = (xfs_log_item_t *)ip0->i_itemp;
603 if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
604 if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(lock_mode, 1))) {
605 xfs_iunlock(ip0, lock_mode);
606 if ((++attempts % 5) == 0)
607 delay(1); /* Don't just spin the CPU */
608 goto again;
609 }
610 } else {
611 xfs_ilock(ip1, xfs_lock_inumorder(lock_mode, 1));
612 }
613}
614
615
616void
617__xfs_iflock(
618 struct xfs_inode *ip)
619{
620 wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_IFLOCK_BIT);
621 DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IFLOCK_BIT);
622
623 do {
624 prepare_to_wait_exclusive(wq, &wait.wait, TASK_UNINTERRUPTIBLE);
625 if (xfs_isiflocked(ip))
626 io_schedule();
627 } while (!xfs_iflock_nowait(ip));
628
629 finish_wait(wq, &wait.wait);
630}
631
632STATIC uint
633_xfs_dic2xflags(
634 __uint16_t di_flags,
635 uint64_t di_flags2,
636 bool has_attr)
637{
638 uint flags = 0;
639
640 if (di_flags & XFS_DIFLAG_ANY) {
641 if (di_flags & XFS_DIFLAG_REALTIME)
642 flags |= FS_XFLAG_REALTIME;
643 if (di_flags & XFS_DIFLAG_PREALLOC)
644 flags |= FS_XFLAG_PREALLOC;
645 if (di_flags & XFS_DIFLAG_IMMUTABLE)
646 flags |= FS_XFLAG_IMMUTABLE;
647 if (di_flags & XFS_DIFLAG_APPEND)
648 flags |= FS_XFLAG_APPEND;
649 if (di_flags & XFS_DIFLAG_SYNC)
650 flags |= FS_XFLAG_SYNC;
651 if (di_flags & XFS_DIFLAG_NOATIME)
652 flags |= FS_XFLAG_NOATIME;
653 if (di_flags & XFS_DIFLAG_NODUMP)
654 flags |= FS_XFLAG_NODUMP;
655 if (di_flags & XFS_DIFLAG_RTINHERIT)
656 flags |= FS_XFLAG_RTINHERIT;
657 if (di_flags & XFS_DIFLAG_PROJINHERIT)
658 flags |= FS_XFLAG_PROJINHERIT;
659 if (di_flags & XFS_DIFLAG_NOSYMLINKS)
660 flags |= FS_XFLAG_NOSYMLINKS;
661 if (di_flags & XFS_DIFLAG_EXTSIZE)
662 flags |= FS_XFLAG_EXTSIZE;
663 if (di_flags & XFS_DIFLAG_EXTSZINHERIT)
664 flags |= FS_XFLAG_EXTSZINHERIT;
665 if (di_flags & XFS_DIFLAG_NODEFRAG)
666 flags |= FS_XFLAG_NODEFRAG;
667 if (di_flags & XFS_DIFLAG_FILESTREAM)
668 flags |= FS_XFLAG_FILESTREAM;
669 }
670
671 if (di_flags2 & XFS_DIFLAG2_ANY) {
672 if (di_flags2 & XFS_DIFLAG2_DAX)
673 flags |= FS_XFLAG_DAX;
674 if (di_flags2 & XFS_DIFLAG2_COWEXTSIZE)
675 flags |= FS_XFLAG_COWEXTSIZE;
676 }
677
678 if (has_attr)
679 flags |= FS_XFLAG_HASATTR;
680
681 return flags;
682}
683
684uint
685xfs_ip2xflags(
686 struct xfs_inode *ip)
687{
688 struct xfs_icdinode *dic = &ip->i_d;
689
690 return _xfs_dic2xflags(dic->di_flags, dic->di_flags2, XFS_IFORK_Q(ip));
691}
692
693/*
694 * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
695 * is allowed, otherwise it has to be an exact match. If a CI match is found,
696 * ci_name->name will point to a the actual name (caller must free) or
697 * will be set to NULL if an exact match is found.
698 */
699int
700xfs_lookup(
701 xfs_inode_t *dp,
702 struct xfs_name *name,
703 xfs_inode_t **ipp,
704 struct xfs_name *ci_name)
705{
706 xfs_ino_t inum;
707 int error;
708
709 trace_xfs_lookup(dp, name);
710
711 if (XFS_FORCED_SHUTDOWN(dp->i_mount))
712 return -EIO;
713
714 error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
715 if (error)
716 goto out_unlock;
717
718 error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp);
719 if (error)
720 goto out_free_name;
721
722 return 0;
723
724out_free_name:
725 if (ci_name)
726 kmem_free(ci_name->name);
727out_unlock:
728 *ipp = NULL;
729 return error;
730}
731
732/*
733 * Allocate an inode on disk and return a copy of its in-core version.
734 * The in-core inode is locked exclusively. Set mode, nlink, and rdev
735 * appropriately within the inode. The uid and gid for the inode are
736 * set according to the contents of the given cred structure.
737 *
738 * Use xfs_dialloc() to allocate the on-disk inode. If xfs_dialloc()
739 * has a free inode available, call xfs_iget() to obtain the in-core
740 * version of the allocated inode. Finally, fill in the inode and
741 * log its initial contents. In this case, ialloc_context would be
742 * set to NULL.
743 *
744 * If xfs_dialloc() does not have an available inode, it will replenish
745 * its supply by doing an allocation. Since we can only do one
746 * allocation within a transaction without deadlocks, we must commit
747 * the current transaction before returning the inode itself.
748 * In this case, therefore, we will set ialloc_context and return.
749 * The caller should then commit the current transaction, start a new
750 * transaction, and call xfs_ialloc() again to actually get the inode.
751 *
752 * To ensure that some other process does not grab the inode that
753 * was allocated during the first call to xfs_ialloc(), this routine
754 * also returns the [locked] bp pointing to the head of the freelist
755 * as ialloc_context. The caller should hold this buffer across
756 * the commit and pass it back into this routine on the second call.
757 *
758 * If we are allocating quota inodes, we do not have a parent inode
759 * to attach to or associate with (i.e. pip == NULL) because they
760 * are not linked into the directory structure - they are attached
761 * directly to the superblock - and so have no parent.
762 */
763static int
764xfs_ialloc(
765 xfs_trans_t *tp,
766 xfs_inode_t *pip,
767 umode_t mode,
768 xfs_nlink_t nlink,
769 xfs_dev_t rdev,
770 prid_t prid,
771 int okalloc,
772 xfs_buf_t **ialloc_context,
773 xfs_inode_t **ipp)
774{
775 struct xfs_mount *mp = tp->t_mountp;
776 xfs_ino_t ino;
777 xfs_inode_t *ip;
778 uint flags;
779 int error;
780 struct timespec tv;
781 struct inode *inode;
782
783 /*
784 * Call the space management code to pick
785 * the on-disk inode to be allocated.
786 */
787 error = xfs_dialloc(tp, pip ? pip->i_ino : 0, mode, okalloc,
788 ialloc_context, &ino);
789 if (error)
790 return error;
791 if (*ialloc_context || ino == NULLFSINO) {
792 *ipp = NULL;
793 return 0;
794 }
795 ASSERT(*ialloc_context == NULL);
796
797 /*
798 * Get the in-core inode with the lock held exclusively.
799 * This is because we're setting fields here we need
800 * to prevent others from looking at until we're done.
801 */
802 error = xfs_iget(mp, tp, ino, XFS_IGET_CREATE,
803 XFS_ILOCK_EXCL, &ip);
804 if (error)
805 return error;
806 ASSERT(ip != NULL);
807 inode = VFS_I(ip);
808
809 /*
810 * We always convert v1 inodes to v2 now - we only support filesystems
811 * with >= v2 inode capability, so there is no reason for ever leaving
812 * an inode in v1 format.
813 */
814 if (ip->i_d.di_version == 1)
815 ip->i_d.di_version = 2;
816
817 inode->i_mode = mode;
818 set_nlink(inode, nlink);
819 ip->i_d.di_uid = xfs_kuid_to_uid(current_fsuid());
820 ip->i_d.di_gid = xfs_kgid_to_gid(current_fsgid());
821 xfs_set_projid(ip, prid);
822
823 if (pip && XFS_INHERIT_GID(pip)) {
824 ip->i_d.di_gid = pip->i_d.di_gid;
825 if ((VFS_I(pip)->i_mode & S_ISGID) && S_ISDIR(mode))
826 inode->i_mode |= S_ISGID;
827 }
828
829 /*
830 * If the group ID of the new file does not match the effective group
831 * ID or one of the supplementary group IDs, the S_ISGID bit is cleared
832 * (and only if the irix_sgid_inherit compatibility variable is set).
833 */
834 if ((irix_sgid_inherit) &&
835 (inode->i_mode & S_ISGID) &&
836 (!in_group_p(xfs_gid_to_kgid(ip->i_d.di_gid))))
837 inode->i_mode &= ~S_ISGID;
838
839 ip->i_d.di_size = 0;
840 ip->i_d.di_nextents = 0;
841 ASSERT(ip->i_d.di_nblocks == 0);
842
843 tv = current_time(inode);
844 inode->i_mtime = tv;
845 inode->i_atime = tv;
846 inode->i_ctime = tv;
847
848 ip->i_d.di_extsize = 0;
849 ip->i_d.di_dmevmask = 0;
850 ip->i_d.di_dmstate = 0;
851 ip->i_d.di_flags = 0;
852
853 if (ip->i_d.di_version == 3) {
854 inode->i_version = 1;
855 ip->i_d.di_flags2 = 0;
856 ip->i_d.di_cowextsize = 0;
857 ip->i_d.di_crtime.t_sec = (__int32_t)tv.tv_sec;
858 ip->i_d.di_crtime.t_nsec = (__int32_t)tv.tv_nsec;
859 }
860
861
862 flags = XFS_ILOG_CORE;
863 switch (mode & S_IFMT) {
864 case S_IFIFO:
865 case S_IFCHR:
866 case S_IFBLK:
867 case S_IFSOCK:
868 ip->i_d.di_format = XFS_DINODE_FMT_DEV;
869 ip->i_df.if_u2.if_rdev = rdev;
870 ip->i_df.if_flags = 0;
871 flags |= XFS_ILOG_DEV;
872 break;
873 case S_IFREG:
874 case S_IFDIR:
875 if (pip && (pip->i_d.di_flags & XFS_DIFLAG_ANY)) {
876 uint64_t di_flags2 = 0;
877 uint di_flags = 0;
878
879 if (S_ISDIR(mode)) {
880 if (pip->i_d.di_flags & XFS_DIFLAG_RTINHERIT)
881 di_flags |= XFS_DIFLAG_RTINHERIT;
882 if (pip->i_d.di_flags & XFS_DIFLAG_EXTSZINHERIT) {
883 di_flags |= XFS_DIFLAG_EXTSZINHERIT;
884 ip->i_d.di_extsize = pip->i_d.di_extsize;
885 }
886 if (pip->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
887 di_flags |= XFS_DIFLAG_PROJINHERIT;
888 } else if (S_ISREG(mode)) {
889 if (pip->i_d.di_flags & XFS_DIFLAG_RTINHERIT)
890 di_flags |= XFS_DIFLAG_REALTIME;
891 if (pip->i_d.di_flags & XFS_DIFLAG_EXTSZINHERIT) {
892 di_flags |= XFS_DIFLAG_EXTSIZE;
893 ip->i_d.di_extsize = pip->i_d.di_extsize;
894 }
895 }
896 if ((pip->i_d.di_flags & XFS_DIFLAG_NOATIME) &&
897 xfs_inherit_noatime)
898 di_flags |= XFS_DIFLAG_NOATIME;
899 if ((pip->i_d.di_flags & XFS_DIFLAG_NODUMP) &&
900 xfs_inherit_nodump)
901 di_flags |= XFS_DIFLAG_NODUMP;
902 if ((pip->i_d.di_flags & XFS_DIFLAG_SYNC) &&
903 xfs_inherit_sync)
904 di_flags |= XFS_DIFLAG_SYNC;
905 if ((pip->i_d.di_flags & XFS_DIFLAG_NOSYMLINKS) &&
906 xfs_inherit_nosymlinks)
907 di_flags |= XFS_DIFLAG_NOSYMLINKS;
908 if ((pip->i_d.di_flags & XFS_DIFLAG_NODEFRAG) &&
909 xfs_inherit_nodefrag)
910 di_flags |= XFS_DIFLAG_NODEFRAG;
911 if (pip->i_d.di_flags & XFS_DIFLAG_FILESTREAM)
912 di_flags |= XFS_DIFLAG_FILESTREAM;
913 if (pip->i_d.di_flags2 & XFS_DIFLAG2_DAX)
914 di_flags2 |= XFS_DIFLAG2_DAX;
915
916 ip->i_d.di_flags |= di_flags;
917 ip->i_d.di_flags2 |= di_flags2;
918 }
919 if (pip &&
920 (pip->i_d.di_flags2 & XFS_DIFLAG2_ANY) &&
921 pip->i_d.di_version == 3 &&
922 ip->i_d.di_version == 3) {
923 if (pip->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE) {
924 ip->i_d.di_flags2 |= XFS_DIFLAG2_COWEXTSIZE;
925 ip->i_d.di_cowextsize = pip->i_d.di_cowextsize;
926 }
927 }
928 /* FALLTHROUGH */
929 case S_IFLNK:
930 ip->i_d.di_format = XFS_DINODE_FMT_EXTENTS;
931 ip->i_df.if_flags = XFS_IFEXTENTS;
932 ip->i_df.if_bytes = ip->i_df.if_real_bytes = 0;
933 ip->i_df.if_u1.if_extents = NULL;
934 break;
935 default:
936 ASSERT(0);
937 }
938 /*
939 * Attribute fork settings for new inode.
940 */
941 ip->i_d.di_aformat = XFS_DINODE_FMT_EXTENTS;
942 ip->i_d.di_anextents = 0;
943
944 /*
945 * Log the new values stuffed into the inode.
946 */
947 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
948 xfs_trans_log_inode(tp, ip, flags);
949
950 /* now that we have an i_mode we can setup the inode structure */
951 xfs_setup_inode(ip);
952
953 *ipp = ip;
954 return 0;
955}
956
957/*
958 * Allocates a new inode from disk and return a pointer to the
959 * incore copy. This routine will internally commit the current
960 * transaction and allocate a new one if the Space Manager needed
961 * to do an allocation to replenish the inode free-list.
962 *
963 * This routine is designed to be called from xfs_create and
964 * xfs_create_dir.
965 *
966 */
967int
968xfs_dir_ialloc(
969 xfs_trans_t **tpp, /* input: current transaction;
970 output: may be a new transaction. */
971 xfs_inode_t *dp, /* directory within whose allocate
972 the inode. */
973 umode_t mode,
974 xfs_nlink_t nlink,
975 xfs_dev_t rdev,
976 prid_t prid, /* project id */
977 int okalloc, /* ok to allocate new space */
978 xfs_inode_t **ipp, /* pointer to inode; it will be
979 locked. */
980 int *committed)
981
982{
983 xfs_trans_t *tp;
984 xfs_inode_t *ip;
985 xfs_buf_t *ialloc_context = NULL;
986 int code;
987 void *dqinfo;
988 uint tflags;
989
990 tp = *tpp;
991 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
992
993 /*
994 * xfs_ialloc will return a pointer to an incore inode if
995 * the Space Manager has an available inode on the free
996 * list. Otherwise, it will do an allocation and replenish
997 * the freelist. Since we can only do one allocation per
998 * transaction without deadlocks, we will need to commit the
999 * current transaction and start a new one. We will then
1000 * need to call xfs_ialloc again to get the inode.
1001 *
1002 * If xfs_ialloc did an allocation to replenish the freelist,
1003 * it returns the bp containing the head of the freelist as
1004 * ialloc_context. We will hold a lock on it across the
1005 * transaction commit so that no other process can steal
1006 * the inode(s) that we've just allocated.
1007 */
1008 code = xfs_ialloc(tp, dp, mode, nlink, rdev, prid, okalloc,
1009 &ialloc_context, &ip);
1010
1011 /*
1012 * Return an error if we were unable to allocate a new inode.
1013 * This should only happen if we run out of space on disk or
1014 * encounter a disk error.
1015 */
1016 if (code) {
1017 *ipp = NULL;
1018 return code;
1019 }
1020 if (!ialloc_context && !ip) {
1021 *ipp = NULL;
1022 return -ENOSPC;
1023 }
1024
1025 /*
1026 * If the AGI buffer is non-NULL, then we were unable to get an
1027 * inode in one operation. We need to commit the current
1028 * transaction and call xfs_ialloc() again. It is guaranteed
1029 * to succeed the second time.
1030 */
1031 if (ialloc_context) {
1032 /*
1033 * Normally, xfs_trans_commit releases all the locks.
1034 * We call bhold to hang on to the ialloc_context across
1035 * the commit. Holding this buffer prevents any other
1036 * processes from doing any allocations in this
1037 * allocation group.
1038 */
1039 xfs_trans_bhold(tp, ialloc_context);
1040
1041 /*
1042 * We want the quota changes to be associated with the next
1043 * transaction, NOT this one. So, detach the dqinfo from this
1044 * and attach it to the next transaction.
1045 */
1046 dqinfo = NULL;
1047 tflags = 0;
1048 if (tp->t_dqinfo) {
1049 dqinfo = (void *)tp->t_dqinfo;
1050 tp->t_dqinfo = NULL;
1051 tflags = tp->t_flags & XFS_TRANS_DQ_DIRTY;
1052 tp->t_flags &= ~(XFS_TRANS_DQ_DIRTY);
1053 }
1054
1055 code = xfs_trans_roll(&tp, NULL);
1056 if (committed != NULL)
1057 *committed = 1;
1058
1059 /*
1060 * Re-attach the quota info that we detached from prev trx.
1061 */
1062 if (dqinfo) {
1063 tp->t_dqinfo = dqinfo;
1064 tp->t_flags |= tflags;
1065 }
1066
1067 if (code) {
1068 xfs_buf_relse(ialloc_context);
1069 *tpp = tp;
1070 *ipp = NULL;
1071 return code;
1072 }
1073 xfs_trans_bjoin(tp, ialloc_context);
1074
1075 /*
1076 * Call ialloc again. Since we've locked out all
1077 * other allocations in this allocation group,
1078 * this call should always succeed.
1079 */
1080 code = xfs_ialloc(tp, dp, mode, nlink, rdev, prid,
1081 okalloc, &ialloc_context, &ip);
1082
1083 /*
1084 * If we get an error at this point, return to the caller
1085 * so that the current transaction can be aborted.
1086 */
1087 if (code) {
1088 *tpp = tp;
1089 *ipp = NULL;
1090 return code;
1091 }
1092 ASSERT(!ialloc_context && ip);
1093
1094 } else {
1095 if (committed != NULL)
1096 *committed = 0;
1097 }
1098
1099 *ipp = ip;
1100 *tpp = tp;
1101
1102 return 0;
1103}
1104
1105/*
1106 * Decrement the link count on an inode & log the change. If this causes the
1107 * link count to go to zero, move the inode to AGI unlinked list so that it can
1108 * be freed when the last active reference goes away via xfs_inactive().
1109 */
1110static int /* error */
1111xfs_droplink(
1112 xfs_trans_t *tp,
1113 xfs_inode_t *ip)
1114{
1115 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
1116
1117 drop_nlink(VFS_I(ip));
1118 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1119
1120 if (VFS_I(ip)->i_nlink)
1121 return 0;
1122
1123 return xfs_iunlink(tp, ip);
1124}
1125
1126/*
1127 * Increment the link count on an inode & log the change.
1128 */
1129static int
1130xfs_bumplink(
1131 xfs_trans_t *tp,
1132 xfs_inode_t *ip)
1133{
1134 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
1135
1136 ASSERT(ip->i_d.di_version > 1);
1137 inc_nlink(VFS_I(ip));
1138 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1139 return 0;
1140}
1141
1142int
1143xfs_create(
1144 xfs_inode_t *dp,
1145 struct xfs_name *name,
1146 umode_t mode,
1147 xfs_dev_t rdev,
1148 xfs_inode_t **ipp)
1149{
1150 int is_dir = S_ISDIR(mode);
1151 struct xfs_mount *mp = dp->i_mount;
1152 struct xfs_inode *ip = NULL;
1153 struct xfs_trans *tp = NULL;
1154 int error;
1155 struct xfs_defer_ops dfops;
1156 xfs_fsblock_t first_block;
1157 bool unlock_dp_on_error = false;
1158 prid_t prid;
1159 struct xfs_dquot *udqp = NULL;
1160 struct xfs_dquot *gdqp = NULL;
1161 struct xfs_dquot *pdqp = NULL;
1162 struct xfs_trans_res *tres;
1163 uint resblks;
1164
1165 trace_xfs_create(dp, name);
1166
1167 if (XFS_FORCED_SHUTDOWN(mp))
1168 return -EIO;
1169
1170 prid = xfs_get_initial_prid(dp);
1171
1172 /*
1173 * Make sure that we have allocated dquot(s) on disk.
1174 */
1175 error = xfs_qm_vop_dqalloc(dp, xfs_kuid_to_uid(current_fsuid()),
1176 xfs_kgid_to_gid(current_fsgid()), prid,
1177 XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
1178 &udqp, &gdqp, &pdqp);
1179 if (error)
1180 return error;
1181
1182 if (is_dir) {
1183 rdev = 0;
1184 resblks = XFS_MKDIR_SPACE_RES(mp, name->len);
1185 tres = &M_RES(mp)->tr_mkdir;
1186 } else {
1187 resblks = XFS_CREATE_SPACE_RES(mp, name->len);
1188 tres = &M_RES(mp)->tr_create;
1189 }
1190
1191 /*
1192 * Initially assume that the file does not exist and
1193 * reserve the resources for that case. If that is not
1194 * the case we'll drop the one we have and get a more
1195 * appropriate transaction later.
1196 */
1197 error = xfs_trans_alloc(mp, tres, resblks, 0, 0, &tp);
1198 if (error == -ENOSPC) {
1199 /* flush outstanding delalloc blocks and retry */
1200 xfs_flush_inodes(mp);
1201 error = xfs_trans_alloc(mp, tres, resblks, 0, 0, &tp);
1202 }
1203 if (error == -ENOSPC) {
1204 /* No space at all so try a "no-allocation" reservation */
1205 resblks = 0;
1206 error = xfs_trans_alloc(mp, tres, 0, 0, 0, &tp);
1207 }
1208 if (error)
1209 goto out_release_inode;
1210
1211 xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
1212 unlock_dp_on_error = true;
1213
1214 xfs_defer_init(&dfops, &first_block);
1215
1216 /*
1217 * Reserve disk quota and the inode.
1218 */
1219 error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp,
1220 pdqp, resblks, 1, 0);
1221 if (error)
1222 goto out_trans_cancel;
1223
1224 if (!resblks) {
1225 error = xfs_dir_canenter(tp, dp, name);
1226 if (error)
1227 goto out_trans_cancel;
1228 }
1229
1230 /*
1231 * A newly created regular or special file just has one directory
1232 * entry pointing to them, but a directory also the "." entry
1233 * pointing to itself.
1234 */
1235 error = xfs_dir_ialloc(&tp, dp, mode, is_dir ? 2 : 1, rdev,
1236 prid, resblks > 0, &ip, NULL);
1237 if (error)
1238 goto out_trans_cancel;
1239
1240 /*
1241 * Now we join the directory inode to the transaction. We do not do it
1242 * earlier because xfs_dir_ialloc might commit the previous transaction
1243 * (and release all the locks). An error from here on will result in
1244 * the transaction cancel unlocking dp so don't do it explicitly in the
1245 * error path.
1246 */
1247 xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1248 unlock_dp_on_error = false;
1249
1250 error = xfs_dir_createname(tp, dp, name, ip->i_ino,
1251 &first_block, &dfops, resblks ?
1252 resblks - XFS_IALLOC_SPACE_RES(mp) : 0);
1253 if (error) {
1254 ASSERT(error != -ENOSPC);
1255 goto out_trans_cancel;
1256 }
1257 xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1258 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1259
1260 if (is_dir) {
1261 error = xfs_dir_init(tp, ip, dp);
1262 if (error)
1263 goto out_bmap_cancel;
1264
1265 error = xfs_bumplink(tp, dp);
1266 if (error)
1267 goto out_bmap_cancel;
1268 }
1269
1270 /*
1271 * If this is a synchronous mount, make sure that the
1272 * create transaction goes to disk before returning to
1273 * the user.
1274 */
1275 if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1276 xfs_trans_set_sync(tp);
1277
1278 /*
1279 * Attach the dquot(s) to the inodes and modify them incore.
1280 * These ids of the inode couldn't have changed since the new
1281 * inode has been locked ever since it was created.
1282 */
1283 xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1284
1285 error = xfs_defer_finish(&tp, &dfops, NULL);
1286 if (error)
1287 goto out_bmap_cancel;
1288
1289 error = xfs_trans_commit(tp);
1290 if (error)
1291 goto out_release_inode;
1292
1293 xfs_qm_dqrele(udqp);
1294 xfs_qm_dqrele(gdqp);
1295 xfs_qm_dqrele(pdqp);
1296
1297 *ipp = ip;
1298 return 0;
1299
1300 out_bmap_cancel:
1301 xfs_defer_cancel(&dfops);
1302 out_trans_cancel:
1303 xfs_trans_cancel(tp);
1304 out_release_inode:
1305 /*
1306 * Wait until after the current transaction is aborted to finish the
1307 * setup of the inode and release the inode. This prevents recursive
1308 * transactions and deadlocks from xfs_inactive.
1309 */
1310 if (ip) {
1311 xfs_finish_inode_setup(ip);
1312 IRELE(ip);
1313 }
1314
1315 xfs_qm_dqrele(udqp);
1316 xfs_qm_dqrele(gdqp);
1317 xfs_qm_dqrele(pdqp);
1318
1319 if (unlock_dp_on_error)
1320 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1321 return error;
1322}
1323
1324int
1325xfs_create_tmpfile(
1326 struct xfs_inode *dp,
1327 struct dentry *dentry,
1328 umode_t mode,
1329 struct xfs_inode **ipp)
1330{
1331 struct xfs_mount *mp = dp->i_mount;
1332 struct xfs_inode *ip = NULL;
1333 struct xfs_trans *tp = NULL;
1334 int error;
1335 prid_t prid;
1336 struct xfs_dquot *udqp = NULL;
1337 struct xfs_dquot *gdqp = NULL;
1338 struct xfs_dquot *pdqp = NULL;
1339 struct xfs_trans_res *tres;
1340 uint resblks;
1341
1342 if (XFS_FORCED_SHUTDOWN(mp))
1343 return -EIO;
1344
1345 prid = xfs_get_initial_prid(dp);
1346
1347 /*
1348 * Make sure that we have allocated dquot(s) on disk.
1349 */
1350 error = xfs_qm_vop_dqalloc(dp, xfs_kuid_to_uid(current_fsuid()),
1351 xfs_kgid_to_gid(current_fsgid()), prid,
1352 XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
1353 &udqp, &gdqp, &pdqp);
1354 if (error)
1355 return error;
1356
1357 resblks = XFS_IALLOC_SPACE_RES(mp);
1358 tres = &M_RES(mp)->tr_create_tmpfile;
1359
1360 error = xfs_trans_alloc(mp, tres, resblks, 0, 0, &tp);
1361 if (error == -ENOSPC) {
1362 /* No space at all so try a "no-allocation" reservation */
1363 resblks = 0;
1364 error = xfs_trans_alloc(mp, tres, 0, 0, 0, &tp);
1365 }
1366 if (error)
1367 goto out_release_inode;
1368
1369 error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp,
1370 pdqp, resblks, 1, 0);
1371 if (error)
1372 goto out_trans_cancel;
1373
1374 error = xfs_dir_ialloc(&tp, dp, mode, 1, 0,
1375 prid, resblks > 0, &ip, NULL);
1376 if (error)
1377 goto out_trans_cancel;
1378
1379 if (mp->m_flags & XFS_MOUNT_WSYNC)
1380 xfs_trans_set_sync(tp);
1381
1382 /*
1383 * Attach the dquot(s) to the inodes and modify them incore.
1384 * These ids of the inode couldn't have changed since the new
1385 * inode has been locked ever since it was created.
1386 */
1387 xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1388
1389 error = xfs_iunlink(tp, ip);
1390 if (error)
1391 goto out_trans_cancel;
1392
1393 error = xfs_trans_commit(tp);
1394 if (error)
1395 goto out_release_inode;
1396
1397 xfs_qm_dqrele(udqp);
1398 xfs_qm_dqrele(gdqp);
1399 xfs_qm_dqrele(pdqp);
1400
1401 *ipp = ip;
1402 return 0;
1403
1404 out_trans_cancel:
1405 xfs_trans_cancel(tp);
1406 out_release_inode:
1407 /*
1408 * Wait until after the current transaction is aborted to finish the
1409 * setup of the inode and release the inode. This prevents recursive
1410 * transactions and deadlocks from xfs_inactive.
1411 */
1412 if (ip) {
1413 xfs_finish_inode_setup(ip);
1414 IRELE(ip);
1415 }
1416
1417 xfs_qm_dqrele(udqp);
1418 xfs_qm_dqrele(gdqp);
1419 xfs_qm_dqrele(pdqp);
1420
1421 return error;
1422}
1423
1424int
1425xfs_link(
1426 xfs_inode_t *tdp,
1427 xfs_inode_t *sip,
1428 struct xfs_name *target_name)
1429{
1430 xfs_mount_t *mp = tdp->i_mount;
1431 xfs_trans_t *tp;
1432 int error;
1433 struct xfs_defer_ops dfops;
1434 xfs_fsblock_t first_block;
1435 int resblks;
1436
1437 trace_xfs_link(tdp, target_name);
1438
1439 ASSERT(!S_ISDIR(VFS_I(sip)->i_mode));
1440
1441 if (XFS_FORCED_SHUTDOWN(mp))
1442 return -EIO;
1443
1444 error = xfs_qm_dqattach(sip, 0);
1445 if (error)
1446 goto std_return;
1447
1448 error = xfs_qm_dqattach(tdp, 0);
1449 if (error)
1450 goto std_return;
1451
1452 resblks = XFS_LINK_SPACE_RES(mp, target_name->len);
1453 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_link, resblks, 0, 0, &tp);
1454 if (error == -ENOSPC) {
1455 resblks = 0;
1456 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_link, 0, 0, 0, &tp);
1457 }
1458 if (error)
1459 goto std_return;
1460
1461 xfs_lock_two_inodes(sip, tdp, XFS_ILOCK_EXCL);
1462
1463 xfs_trans_ijoin(tp, sip, XFS_ILOCK_EXCL);
1464 xfs_trans_ijoin(tp, tdp, XFS_ILOCK_EXCL);
1465
1466 /*
1467 * If we are using project inheritance, we only allow hard link
1468 * creation in our tree when the project IDs are the same; else
1469 * the tree quota mechanism could be circumvented.
1470 */
1471 if (unlikely((tdp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
1472 (xfs_get_projid(tdp) != xfs_get_projid(sip)))) {
1473 error = -EXDEV;
1474 goto error_return;
1475 }
1476
1477 if (!resblks) {
1478 error = xfs_dir_canenter(tp, tdp, target_name);
1479 if (error)
1480 goto error_return;
1481 }
1482
1483 xfs_defer_init(&dfops, &first_block);
1484
1485 /*
1486 * Handle initial link state of O_TMPFILE inode
1487 */
1488 if (VFS_I(sip)->i_nlink == 0) {
1489 error = xfs_iunlink_remove(tp, sip);
1490 if (error)
1491 goto error_return;
1492 }
1493
1494 error = xfs_dir_createname(tp, tdp, target_name, sip->i_ino,
1495 &first_block, &dfops, resblks);
1496 if (error)
1497 goto error_return;
1498 xfs_trans_ichgtime(tp, tdp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1499 xfs_trans_log_inode(tp, tdp, XFS_ILOG_CORE);
1500
1501 error = xfs_bumplink(tp, sip);
1502 if (error)
1503 goto error_return;
1504
1505 /*
1506 * If this is a synchronous mount, make sure that the
1507 * link transaction goes to disk before returning to
1508 * the user.
1509 */
1510 if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1511 xfs_trans_set_sync(tp);
1512
1513 error = xfs_defer_finish(&tp, &dfops, NULL);
1514 if (error) {
1515 xfs_defer_cancel(&dfops);
1516 goto error_return;
1517 }
1518
1519 return xfs_trans_commit(tp);
1520
1521 error_return:
1522 xfs_trans_cancel(tp);
1523 std_return:
1524 return error;
1525}
1526
1527/*
1528 * Free up the underlying blocks past new_size. The new size must be smaller
1529 * than the current size. This routine can be used both for the attribute and
1530 * data fork, and does not modify the inode size, which is left to the caller.
1531 *
1532 * The transaction passed to this routine must have made a permanent log
1533 * reservation of at least XFS_ITRUNCATE_LOG_RES. This routine may commit the
1534 * given transaction and start new ones, so make sure everything involved in
1535 * the transaction is tidy before calling here. Some transaction will be
1536 * returned to the caller to be committed. The incoming transaction must
1537 * already include the inode, and both inode locks must be held exclusively.
1538 * The inode must also be "held" within the transaction. On return the inode
1539 * will be "held" within the returned transaction. This routine does NOT
1540 * require any disk space to be reserved for it within the transaction.
1541 *
1542 * If we get an error, we must return with the inode locked and linked into the
1543 * current transaction. This keeps things simple for the higher level code,
1544 * because it always knows that the inode is locked and held in the transaction
1545 * that returns to it whether errors occur or not. We don't mark the inode
1546 * dirty on error so that transactions can be easily aborted if possible.
1547 */
1548int
1549xfs_itruncate_extents(
1550 struct xfs_trans **tpp,
1551 struct xfs_inode *ip,
1552 int whichfork,
1553 xfs_fsize_t new_size)
1554{
1555 struct xfs_mount *mp = ip->i_mount;
1556 struct xfs_trans *tp = *tpp;
1557 struct xfs_defer_ops dfops;
1558 xfs_fsblock_t first_block;
1559 xfs_fileoff_t first_unmap_block;
1560 xfs_fileoff_t last_block;
1561 xfs_filblks_t unmap_len;
1562 int error = 0;
1563 int done = 0;
1564
1565 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1566 ASSERT(!atomic_read(&VFS_I(ip)->i_count) ||
1567 xfs_isilocked(ip, XFS_IOLOCK_EXCL));
1568 ASSERT(new_size <= XFS_ISIZE(ip));
1569 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
1570 ASSERT(ip->i_itemp != NULL);
1571 ASSERT(ip->i_itemp->ili_lock_flags == 0);
1572 ASSERT(!XFS_NOT_DQATTACHED(mp, ip));
1573
1574 trace_xfs_itruncate_extents_start(ip, new_size);
1575
1576 /*
1577 * Since it is possible for space to become allocated beyond
1578 * the end of the file (in a crash where the space is allocated
1579 * but the inode size is not yet updated), simply remove any
1580 * blocks which show up between the new EOF and the maximum
1581 * possible file size. If the first block to be removed is
1582 * beyond the maximum file size (ie it is the same as last_block),
1583 * then there is nothing to do.
1584 */
1585 first_unmap_block = XFS_B_TO_FSB(mp, (xfs_ufsize_t)new_size);
1586 last_block = XFS_B_TO_FSB(mp, mp->m_super->s_maxbytes);
1587 if (first_unmap_block == last_block)
1588 return 0;
1589
1590 ASSERT(first_unmap_block < last_block);
1591 unmap_len = last_block - first_unmap_block + 1;
1592 while (!done) {
1593 xfs_defer_init(&dfops, &first_block);
1594 error = xfs_bunmapi(tp, ip,
1595 first_unmap_block, unmap_len,
1596 xfs_bmapi_aflag(whichfork),
1597 XFS_ITRUNC_MAX_EXTENTS,
1598 &first_block, &dfops,
1599 &done);
1600 if (error)
1601 goto out_bmap_cancel;
1602
1603 /*
1604 * Duplicate the transaction that has the permanent
1605 * reservation and commit the old transaction.
1606 */
1607 error = xfs_defer_finish(&tp, &dfops, ip);
1608 if (error)
1609 goto out_bmap_cancel;
1610
1611 error = xfs_trans_roll(&tp, ip);
1612 if (error)
1613 goto out;
1614 }
1615
1616 /* Remove all pending CoW reservations. */
1617 error = xfs_reflink_cancel_cow_blocks(ip, &tp, first_unmap_block,
1618 last_block, true);
1619 if (error)
1620 goto out;
1621
1622 /*
1623 * Clear the reflink flag if we truncated everything.
1624 */
1625 if (ip->i_d.di_nblocks == 0 && xfs_is_reflink_inode(ip)) {
1626 ip->i_d.di_flags2 &= ~XFS_DIFLAG2_REFLINK;
1627 xfs_inode_clear_cowblocks_tag(ip);
1628 }
1629
1630 /*
1631 * Always re-log the inode so that our permanent transaction can keep
1632 * on rolling it forward in the log.
1633 */
1634 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1635
1636 trace_xfs_itruncate_extents_end(ip, new_size);
1637
1638out:
1639 *tpp = tp;
1640 return error;
1641out_bmap_cancel:
1642 /*
1643 * If the bunmapi call encounters an error, return to the caller where
1644 * the transaction can be properly aborted. We just need to make sure
1645 * we're not holding any resources that we were not when we came in.
1646 */
1647 xfs_defer_cancel(&dfops);
1648 goto out;
1649}
1650
1651int
1652xfs_release(
1653 xfs_inode_t *ip)
1654{
1655 xfs_mount_t *mp = ip->i_mount;
1656 int error;
1657
1658 if (!S_ISREG(VFS_I(ip)->i_mode) || (VFS_I(ip)->i_mode == 0))
1659 return 0;
1660
1661 /* If this is a read-only mount, don't do this (would generate I/O) */
1662 if (mp->m_flags & XFS_MOUNT_RDONLY)
1663 return 0;
1664
1665 if (!XFS_FORCED_SHUTDOWN(mp)) {
1666 int truncated;
1667
1668 /*
1669 * If we previously truncated this file and removed old data
1670 * in the process, we want to initiate "early" writeout on
1671 * the last close. This is an attempt to combat the notorious
1672 * NULL files problem which is particularly noticeable from a
1673 * truncate down, buffered (re-)write (delalloc), followed by
1674 * a crash. What we are effectively doing here is
1675 * significantly reducing the time window where we'd otherwise
1676 * be exposed to that problem.
1677 */
1678 truncated = xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED);
1679 if (truncated) {
1680 xfs_iflags_clear(ip, XFS_IDIRTY_RELEASE);
1681 if (ip->i_delayed_blks > 0) {
1682 error = filemap_flush(VFS_I(ip)->i_mapping);
1683 if (error)
1684 return error;
1685 }
1686 }
1687 }
1688
1689 if (VFS_I(ip)->i_nlink == 0)
1690 return 0;
1691
1692 if (xfs_can_free_eofblocks(ip, false)) {
1693
1694 /*
1695 * Check if the inode is being opened, written and closed
1696 * frequently and we have delayed allocation blocks outstanding
1697 * (e.g. streaming writes from the NFS server), truncating the
1698 * blocks past EOF will cause fragmentation to occur.
1699 *
1700 * In this case don't do the truncation, but we have to be
1701 * careful how we detect this case. Blocks beyond EOF show up as
1702 * i_delayed_blks even when the inode is clean, so we need to
1703 * truncate them away first before checking for a dirty release.
1704 * Hence on the first dirty close we will still remove the
1705 * speculative allocation, but after that we will leave it in
1706 * place.
1707 */
1708 if (xfs_iflags_test(ip, XFS_IDIRTY_RELEASE))
1709 return 0;
1710 /*
1711 * If we can't get the iolock just skip truncating the blocks
1712 * past EOF because we could deadlock with the mmap_sem
1713 * otherwise. We'll get another chance to drop them once the
1714 * last reference to the inode is dropped, so we'll never leak
1715 * blocks permanently.
1716 */
1717 if (xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1718 error = xfs_free_eofblocks(ip);
1719 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1720 if (error)
1721 return error;
1722 }
1723
1724 /* delalloc blocks after truncation means it really is dirty */
1725 if (ip->i_delayed_blks)
1726 xfs_iflags_set(ip, XFS_IDIRTY_RELEASE);
1727 }
1728 return 0;
1729}
1730
1731/*
1732 * xfs_inactive_truncate
1733 *
1734 * Called to perform a truncate when an inode becomes unlinked.
1735 */
1736STATIC int
1737xfs_inactive_truncate(
1738 struct xfs_inode *ip)
1739{
1740 struct xfs_mount *mp = ip->i_mount;
1741 struct xfs_trans *tp;
1742 int error;
1743
1744 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate, 0, 0, 0, &tp);
1745 if (error) {
1746 ASSERT(XFS_FORCED_SHUTDOWN(mp));
1747 return error;
1748 }
1749
1750 xfs_ilock(ip, XFS_ILOCK_EXCL);
1751 xfs_trans_ijoin(tp, ip, 0);
1752
1753 /*
1754 * Log the inode size first to prevent stale data exposure in the event
1755 * of a system crash before the truncate completes. See the related
1756 * comment in xfs_vn_setattr_size() for details.
1757 */
1758 ip->i_d.di_size = 0;
1759 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1760
1761 error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK, 0);
1762 if (error)
1763 goto error_trans_cancel;
1764
1765 ASSERT(ip->i_d.di_nextents == 0);
1766
1767 error = xfs_trans_commit(tp);
1768 if (error)
1769 goto error_unlock;
1770
1771 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1772 return 0;
1773
1774error_trans_cancel:
1775 xfs_trans_cancel(tp);
1776error_unlock:
1777 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1778 return error;
1779}
1780
1781/*
1782 * xfs_inactive_ifree()
1783 *
1784 * Perform the inode free when an inode is unlinked.
1785 */
1786STATIC int
1787xfs_inactive_ifree(
1788 struct xfs_inode *ip)
1789{
1790 struct xfs_defer_ops dfops;
1791 xfs_fsblock_t first_block;
1792 struct xfs_mount *mp = ip->i_mount;
1793 struct xfs_trans *tp;
1794 int error;
1795
1796 /*
1797 * We try to use a per-AG reservation for any block needed by the finobt
1798 * tree, but as the finobt feature predates the per-AG reservation
1799 * support a degraded file system might not have enough space for the
1800 * reservation at mount time. In that case try to dip into the reserved
1801 * pool and pray.
1802 *
1803 * Send a warning if the reservation does happen to fail, as the inode
1804 * now remains allocated and sits on the unlinked list until the fs is
1805 * repaired.
1806 */
1807 if (unlikely(mp->m_inotbt_nores)) {
1808 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree,
1809 XFS_IFREE_SPACE_RES(mp), 0, XFS_TRANS_RESERVE,
1810 &tp);
1811 } else {
1812 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree, 0, 0, 0, &tp);
1813 }
1814 if (error) {
1815 if (error == -ENOSPC) {
1816 xfs_warn_ratelimited(mp,
1817 "Failed to remove inode(s) from unlinked list. "
1818 "Please free space, unmount and run xfs_repair.");
1819 } else {
1820 ASSERT(XFS_FORCED_SHUTDOWN(mp));
1821 }
1822 return error;
1823 }
1824
1825 xfs_ilock(ip, XFS_ILOCK_EXCL);
1826 xfs_trans_ijoin(tp, ip, 0);
1827
1828 xfs_defer_init(&dfops, &first_block);
1829 error = xfs_ifree(tp, ip, &dfops);
1830 if (error) {
1831 /*
1832 * If we fail to free the inode, shut down. The cancel
1833 * might do that, we need to make sure. Otherwise the
1834 * inode might be lost for a long time or forever.
1835 */
1836 if (!XFS_FORCED_SHUTDOWN(mp)) {
1837 xfs_notice(mp, "%s: xfs_ifree returned error %d",
1838 __func__, error);
1839 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1840 }
1841 xfs_trans_cancel(tp);
1842 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1843 return error;
1844 }
1845
1846 /*
1847 * Credit the quota account(s). The inode is gone.
1848 */
1849 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
1850
1851 /*
1852 * Just ignore errors at this point. There is nothing we can do except
1853 * to try to keep going. Make sure it's not a silent error.
1854 */
1855 error = xfs_defer_finish(&tp, &dfops, NULL);
1856 if (error) {
1857 xfs_notice(mp, "%s: xfs_defer_finish returned error %d",
1858 __func__, error);
1859 xfs_defer_cancel(&dfops);
1860 }
1861 error = xfs_trans_commit(tp);
1862 if (error)
1863 xfs_notice(mp, "%s: xfs_trans_commit returned error %d",
1864 __func__, error);
1865
1866 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1867 return 0;
1868}
1869
1870/*
1871 * xfs_inactive
1872 *
1873 * This is called when the vnode reference count for the vnode
1874 * goes to zero. If the file has been unlinked, then it must
1875 * now be truncated. Also, we clear all of the read-ahead state
1876 * kept for the inode here since the file is now closed.
1877 */
1878void
1879xfs_inactive(
1880 xfs_inode_t *ip)
1881{
1882 struct xfs_mount *mp;
1883 int error;
1884 int truncate = 0;
1885
1886 /*
1887 * If the inode is already free, then there can be nothing
1888 * to clean up here.
1889 */
1890 if (VFS_I(ip)->i_mode == 0) {
1891 ASSERT(ip->i_df.if_real_bytes == 0);
1892 ASSERT(ip->i_df.if_broot_bytes == 0);
1893 return;
1894 }
1895
1896 mp = ip->i_mount;
1897 ASSERT(!xfs_iflags_test(ip, XFS_IRECOVERY));
1898
1899 /* If this is a read-only mount, don't do this (would generate I/O) */
1900 if (mp->m_flags & XFS_MOUNT_RDONLY)
1901 return;
1902
1903 if (VFS_I(ip)->i_nlink != 0) {
1904 /*
1905 * force is true because we are evicting an inode from the
1906 * cache. Post-eof blocks must be freed, lest we end up with
1907 * broken free space accounting.
1908 */
1909 if (xfs_can_free_eofblocks(ip, true)) {
1910 xfs_ilock(ip, XFS_IOLOCK_EXCL);
1911 xfs_free_eofblocks(ip);
1912 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1913 }
1914
1915 return;
1916 }
1917
1918 if (S_ISREG(VFS_I(ip)->i_mode) &&
1919 (ip->i_d.di_size != 0 || XFS_ISIZE(ip) != 0 ||
1920 ip->i_d.di_nextents > 0 || ip->i_delayed_blks > 0))
1921 truncate = 1;
1922
1923 error = xfs_qm_dqattach(ip, 0);
1924 if (error)
1925 return;
1926
1927 if (S_ISLNK(VFS_I(ip)->i_mode))
1928 error = xfs_inactive_symlink(ip);
1929 else if (truncate)
1930 error = xfs_inactive_truncate(ip);
1931 if (error)
1932 return;
1933
1934 /*
1935 * If there are attributes associated with the file then blow them away
1936 * now. The code calls a routine that recursively deconstructs the
1937 * attribute fork. If also blows away the in-core attribute fork.
1938 */
1939 if (XFS_IFORK_Q(ip)) {
1940 error = xfs_attr_inactive(ip);
1941 if (error)
1942 return;
1943 }
1944
1945 ASSERT(!ip->i_afp);
1946 ASSERT(ip->i_d.di_anextents == 0);
1947 ASSERT(ip->i_d.di_forkoff == 0);
1948
1949 /*
1950 * Free the inode.
1951 */
1952 error = xfs_inactive_ifree(ip);
1953 if (error)
1954 return;
1955
1956 /*
1957 * Release the dquots held by inode, if any.
1958 */
1959 xfs_qm_dqdetach(ip);
1960}
1961
1962/*
1963 * This is called when the inode's link count goes to 0 or we are creating a
1964 * tmpfile via O_TMPFILE. In the case of a tmpfile, @ignore_linkcount will be
1965 * set to true as the link count is dropped to zero by the VFS after we've
1966 * created the file successfully, so we have to add it to the unlinked list
1967 * while the link count is non-zero.
1968 *
1969 * We place the on-disk inode on a list in the AGI. It will be pulled from this
1970 * list when the inode is freed.
1971 */
1972STATIC int
1973xfs_iunlink(
1974 struct xfs_trans *tp,
1975 struct xfs_inode *ip)
1976{
1977 xfs_mount_t *mp = tp->t_mountp;
1978 xfs_agi_t *agi;
1979 xfs_dinode_t *dip;
1980 xfs_buf_t *agibp;
1981 xfs_buf_t *ibp;
1982 xfs_agino_t agino;
1983 short bucket_index;
1984 int offset;
1985 int error;
1986
1987 ASSERT(VFS_I(ip)->i_mode != 0);
1988
1989 /*
1990 * Get the agi buffer first. It ensures lock ordering
1991 * on the list.
1992 */
1993 error = xfs_read_agi(mp, tp, XFS_INO_TO_AGNO(mp, ip->i_ino), &agibp);
1994 if (error)
1995 return error;
1996 agi = XFS_BUF_TO_AGI(agibp);
1997
1998 /*
1999 * Get the index into the agi hash table for the
2000 * list this inode will go on.
2001 */
2002 agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2003 ASSERT(agino != 0);
2004 bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
2005 ASSERT(agi->agi_unlinked[bucket_index]);
2006 ASSERT(be32_to_cpu(agi->agi_unlinked[bucket_index]) != agino);
2007
2008 if (agi->agi_unlinked[bucket_index] != cpu_to_be32(NULLAGINO)) {
2009 /*
2010 * There is already another inode in the bucket we need
2011 * to add ourselves to. Add us at the front of the list.
2012 * Here we put the head pointer into our next pointer,
2013 * and then we fall through to point the head at us.
2014 */
2015 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &dip, &ibp,
2016 0, 0);
2017 if (error)
2018 return error;
2019
2020 ASSERT(dip->di_next_unlinked == cpu_to_be32(NULLAGINO));
2021 dip->di_next_unlinked = agi->agi_unlinked[bucket_index];
2022 offset = ip->i_imap.im_boffset +
2023 offsetof(xfs_dinode_t, di_next_unlinked);
2024
2025 /* need to recalc the inode CRC if appropriate */
2026 xfs_dinode_calc_crc(mp, dip);
2027
2028 xfs_trans_inode_buf(tp, ibp);
2029 xfs_trans_log_buf(tp, ibp, offset,
2030 (offset + sizeof(xfs_agino_t) - 1));
2031 xfs_inobp_check(mp, ibp);
2032 }
2033
2034 /*
2035 * Point the bucket head pointer at the inode being inserted.
2036 */
2037 ASSERT(agino != 0);
2038 agi->agi_unlinked[bucket_index] = cpu_to_be32(agino);
2039 offset = offsetof(xfs_agi_t, agi_unlinked) +
2040 (sizeof(xfs_agino_t) * bucket_index);
2041 xfs_trans_log_buf(tp, agibp, offset,
2042 (offset + sizeof(xfs_agino_t) - 1));
2043 return 0;
2044}
2045
2046/*
2047 * Pull the on-disk inode from the AGI unlinked list.
2048 */
2049STATIC int
2050xfs_iunlink_remove(
2051 xfs_trans_t *tp,
2052 xfs_inode_t *ip)
2053{
2054 xfs_ino_t next_ino;
2055 xfs_mount_t *mp;
2056 xfs_agi_t *agi;
2057 xfs_dinode_t *dip;
2058 xfs_buf_t *agibp;
2059 xfs_buf_t *ibp;
2060 xfs_agnumber_t agno;
2061 xfs_agino_t agino;
2062 xfs_agino_t next_agino;
2063 xfs_buf_t *last_ibp;
2064 xfs_dinode_t *last_dip = NULL;
2065 short bucket_index;
2066 int offset, last_offset = 0;
2067 int error;
2068
2069 mp = tp->t_mountp;
2070 agno = XFS_INO_TO_AGNO(mp, ip->i_ino);
2071
2072 /*
2073 * Get the agi buffer first. It ensures lock ordering
2074 * on the list.
2075 */
2076 error = xfs_read_agi(mp, tp, agno, &agibp);
2077 if (error)
2078 return error;
2079
2080 agi = XFS_BUF_TO_AGI(agibp);
2081
2082 /*
2083 * Get the index into the agi hash table for the
2084 * list this inode will go on.
2085 */
2086 agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2087 ASSERT(agino != 0);
2088 bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
2089 ASSERT(agi->agi_unlinked[bucket_index] != cpu_to_be32(NULLAGINO));
2090 ASSERT(agi->agi_unlinked[bucket_index]);
2091
2092 if (be32_to_cpu(agi->agi_unlinked[bucket_index]) == agino) {
2093 /*
2094 * We're at the head of the list. Get the inode's on-disk
2095 * buffer to see if there is anyone after us on the list.
2096 * Only modify our next pointer if it is not already NULLAGINO.
2097 * This saves us the overhead of dealing with the buffer when
2098 * there is no need to change it.
2099 */
2100 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &dip, &ibp,
2101 0, 0);
2102 if (error) {
2103 xfs_warn(mp, "%s: xfs_imap_to_bp returned error %d.",
2104 __func__, error);
2105 return error;
2106 }
2107 next_agino = be32_to_cpu(dip->di_next_unlinked);
2108 ASSERT(next_agino != 0);
2109 if (next_agino != NULLAGINO) {
2110 dip->di_next_unlinked = cpu_to_be32(NULLAGINO);
2111 offset = ip->i_imap.im_boffset +
2112 offsetof(xfs_dinode_t, di_next_unlinked);
2113
2114 /* need to recalc the inode CRC if appropriate */
2115 xfs_dinode_calc_crc(mp, dip);
2116
2117 xfs_trans_inode_buf(tp, ibp);
2118 xfs_trans_log_buf(tp, ibp, offset,
2119 (offset + sizeof(xfs_agino_t) - 1));
2120 xfs_inobp_check(mp, ibp);
2121 } else {
2122 xfs_trans_brelse(tp, ibp);
2123 }
2124 /*
2125 * Point the bucket head pointer at the next inode.
2126 */
2127 ASSERT(next_agino != 0);
2128 ASSERT(next_agino != agino);
2129 agi->agi_unlinked[bucket_index] = cpu_to_be32(next_agino);
2130 offset = offsetof(xfs_agi_t, agi_unlinked) +
2131 (sizeof(xfs_agino_t) * bucket_index);
2132 xfs_trans_log_buf(tp, agibp, offset,
2133 (offset + sizeof(xfs_agino_t) - 1));
2134 } else {
2135 /*
2136 * We need to search the list for the inode being freed.
2137 */
2138 next_agino = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2139 last_ibp = NULL;
2140 while (next_agino != agino) {
2141 struct xfs_imap imap;
2142
2143 if (last_ibp)
2144 xfs_trans_brelse(tp, last_ibp);
2145
2146 imap.im_blkno = 0;
2147 next_ino = XFS_AGINO_TO_INO(mp, agno, next_agino);
2148
2149 error = xfs_imap(mp, tp, next_ino, &imap, 0);
2150 if (error) {
2151 xfs_warn(mp,
2152 "%s: xfs_imap returned error %d.",
2153 __func__, error);
2154 return error;
2155 }
2156
2157 error = xfs_imap_to_bp(mp, tp, &imap, &last_dip,
2158 &last_ibp, 0, 0);
2159 if (error) {
2160 xfs_warn(mp,
2161 "%s: xfs_imap_to_bp returned error %d.",
2162 __func__, error);
2163 return error;
2164 }
2165
2166 last_offset = imap.im_boffset;
2167 next_agino = be32_to_cpu(last_dip->di_next_unlinked);
2168 ASSERT(next_agino != NULLAGINO);
2169 ASSERT(next_agino != 0);
2170 }
2171
2172 /*
2173 * Now last_ibp points to the buffer previous to us on the
2174 * unlinked list. Pull us from the list.
2175 */
2176 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &dip, &ibp,
2177 0, 0);
2178 if (error) {
2179 xfs_warn(mp, "%s: xfs_imap_to_bp(2) returned error %d.",
2180 __func__, error);
2181 return error;
2182 }
2183 next_agino = be32_to_cpu(dip->di_next_unlinked);
2184 ASSERT(next_agino != 0);
2185 ASSERT(next_agino != agino);
2186 if (next_agino != NULLAGINO) {
2187 dip->di_next_unlinked = cpu_to_be32(NULLAGINO);
2188 offset = ip->i_imap.im_boffset +
2189 offsetof(xfs_dinode_t, di_next_unlinked);
2190
2191 /* need to recalc the inode CRC if appropriate */
2192 xfs_dinode_calc_crc(mp, dip);
2193
2194 xfs_trans_inode_buf(tp, ibp);
2195 xfs_trans_log_buf(tp, ibp, offset,
2196 (offset + sizeof(xfs_agino_t) - 1));
2197 xfs_inobp_check(mp, ibp);
2198 } else {
2199 xfs_trans_brelse(tp, ibp);
2200 }
2201 /*
2202 * Point the previous inode on the list to the next inode.
2203 */
2204 last_dip->di_next_unlinked = cpu_to_be32(next_agino);
2205 ASSERT(next_agino != 0);
2206 offset = last_offset + offsetof(xfs_dinode_t, di_next_unlinked);
2207
2208 /* need to recalc the inode CRC if appropriate */
2209 xfs_dinode_calc_crc(mp, last_dip);
2210
2211 xfs_trans_inode_buf(tp, last_ibp);
2212 xfs_trans_log_buf(tp, last_ibp, offset,
2213 (offset + sizeof(xfs_agino_t) - 1));
2214 xfs_inobp_check(mp, last_ibp);
2215 }
2216 return 0;
2217}
2218
2219/*
2220 * A big issue when freeing the inode cluster is that we _cannot_ skip any
2221 * inodes that are in memory - they all must be marked stale and attached to
2222 * the cluster buffer.
2223 */
2224STATIC int
2225xfs_ifree_cluster(
2226 xfs_inode_t *free_ip,
2227 xfs_trans_t *tp,
2228 struct xfs_icluster *xic)
2229{
2230 xfs_mount_t *mp = free_ip->i_mount;
2231 int blks_per_cluster;
2232 int inodes_per_cluster;
2233 int nbufs;
2234 int i, j;
2235 int ioffset;
2236 xfs_daddr_t blkno;
2237 xfs_buf_t *bp;
2238 xfs_inode_t *ip;
2239 xfs_inode_log_item_t *iip;
2240 xfs_log_item_t *lip;
2241 struct xfs_perag *pag;
2242 xfs_ino_t inum;
2243
2244 inum = xic->first_ino;
2245 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, inum));
2246 blks_per_cluster = xfs_icluster_size_fsb(mp);
2247 inodes_per_cluster = blks_per_cluster << mp->m_sb.sb_inopblog;
2248 nbufs = mp->m_ialloc_blks / blks_per_cluster;
2249
2250 for (j = 0; j < nbufs; j++, inum += inodes_per_cluster) {
2251 /*
2252 * The allocation bitmap tells us which inodes of the chunk were
2253 * physically allocated. Skip the cluster if an inode falls into
2254 * a sparse region.
2255 */
2256 ioffset = inum - xic->first_ino;
2257 if ((xic->alloc & XFS_INOBT_MASK(ioffset)) == 0) {
2258 ASSERT(do_mod(ioffset, inodes_per_cluster) == 0);
2259 continue;
2260 }
2261
2262 blkno = XFS_AGB_TO_DADDR(mp, XFS_INO_TO_AGNO(mp, inum),
2263 XFS_INO_TO_AGBNO(mp, inum));
2264
2265 /*
2266 * We obtain and lock the backing buffer first in the process
2267 * here, as we have to ensure that any dirty inode that we
2268 * can't get the flush lock on is attached to the buffer.
2269 * If we scan the in-memory inodes first, then buffer IO can
2270 * complete before we get a lock on it, and hence we may fail
2271 * to mark all the active inodes on the buffer stale.
2272 */
2273 bp = xfs_trans_get_buf(tp, mp->m_ddev_targp, blkno,
2274 mp->m_bsize * blks_per_cluster,
2275 XBF_UNMAPPED);
2276
2277 if (!bp)
2278 return -ENOMEM;
2279
2280 /*
2281 * This buffer may not have been correctly initialised as we
2282 * didn't read it from disk. That's not important because we are
2283 * only using to mark the buffer as stale in the log, and to
2284 * attach stale cached inodes on it. That means it will never be
2285 * dispatched for IO. If it is, we want to know about it, and we
2286 * want it to fail. We can acheive this by adding a write
2287 * verifier to the buffer.
2288 */
2289 bp->b_ops = &xfs_inode_buf_ops;
2290
2291 /*
2292 * Walk the inodes already attached to the buffer and mark them
2293 * stale. These will all have the flush locks held, so an
2294 * in-memory inode walk can't lock them. By marking them all
2295 * stale first, we will not attempt to lock them in the loop
2296 * below as the XFS_ISTALE flag will be set.
2297 */
2298 lip = bp->b_fspriv;
2299 while (lip) {
2300 if (lip->li_type == XFS_LI_INODE) {
2301 iip = (xfs_inode_log_item_t *)lip;
2302 ASSERT(iip->ili_logged == 1);
2303 lip->li_cb = xfs_istale_done;
2304 xfs_trans_ail_copy_lsn(mp->m_ail,
2305 &iip->ili_flush_lsn,
2306 &iip->ili_item.li_lsn);
2307 xfs_iflags_set(iip->ili_inode, XFS_ISTALE);
2308 }
2309 lip = lip->li_bio_list;
2310 }
2311
2312
2313 /*
2314 * For each inode in memory attempt to add it to the inode
2315 * buffer and set it up for being staled on buffer IO
2316 * completion. This is safe as we've locked out tail pushing
2317 * and flushing by locking the buffer.
2318 *
2319 * We have already marked every inode that was part of a
2320 * transaction stale above, which means there is no point in
2321 * even trying to lock them.
2322 */
2323 for (i = 0; i < inodes_per_cluster; i++) {
2324retry:
2325 rcu_read_lock();
2326 ip = radix_tree_lookup(&pag->pag_ici_root,
2327 XFS_INO_TO_AGINO(mp, (inum + i)));
2328
2329 /* Inode not in memory, nothing to do */
2330 if (!ip) {
2331 rcu_read_unlock();
2332 continue;
2333 }
2334
2335 /*
2336 * because this is an RCU protected lookup, we could
2337 * find a recently freed or even reallocated inode
2338 * during the lookup. We need to check under the
2339 * i_flags_lock for a valid inode here. Skip it if it
2340 * is not valid, the wrong inode or stale.
2341 */
2342 spin_lock(&ip->i_flags_lock);
2343 if (ip->i_ino != inum + i ||
2344 __xfs_iflags_test(ip, XFS_ISTALE)) {
2345 spin_unlock(&ip->i_flags_lock);
2346 rcu_read_unlock();
2347 continue;
2348 }
2349 spin_unlock(&ip->i_flags_lock);
2350
2351 /*
2352 * Don't try to lock/unlock the current inode, but we
2353 * _cannot_ skip the other inodes that we did not find
2354 * in the list attached to the buffer and are not
2355 * already marked stale. If we can't lock it, back off
2356 * and retry.
2357 */
2358 if (ip != free_ip &&
2359 !xfs_ilock_nowait(ip, XFS_ILOCK_EXCL)) {
2360 rcu_read_unlock();
2361 delay(1);
2362 goto retry;
2363 }
2364 rcu_read_unlock();
2365
2366 xfs_iflock(ip);
2367 xfs_iflags_set(ip, XFS_ISTALE);
2368
2369 /*
2370 * we don't need to attach clean inodes or those only
2371 * with unlogged changes (which we throw away, anyway).
2372 */
2373 iip = ip->i_itemp;
2374 if (!iip || xfs_inode_clean(ip)) {
2375 ASSERT(ip != free_ip);
2376 xfs_ifunlock(ip);
2377 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2378 continue;
2379 }
2380
2381 iip->ili_last_fields = iip->ili_fields;
2382 iip->ili_fields = 0;
2383 iip->ili_fsync_fields = 0;
2384 iip->ili_logged = 1;
2385 xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
2386 &iip->ili_item.li_lsn);
2387
2388 xfs_buf_attach_iodone(bp, xfs_istale_done,
2389 &iip->ili_item);
2390
2391 if (ip != free_ip)
2392 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2393 }
2394
2395 xfs_trans_stale_inode_buf(tp, bp);
2396 xfs_trans_binval(tp, bp);
2397 }
2398
2399 xfs_perag_put(pag);
2400 return 0;
2401}
2402
2403/*
2404 * This is called to return an inode to the inode free list.
2405 * The inode should already be truncated to 0 length and have
2406 * no pages associated with it. This routine also assumes that
2407 * the inode is already a part of the transaction.
2408 *
2409 * The on-disk copy of the inode will have been added to the list
2410 * of unlinked inodes in the AGI. We need to remove the inode from
2411 * that list atomically with respect to freeing it here.
2412 */
2413int
2414xfs_ifree(
2415 xfs_trans_t *tp,
2416 xfs_inode_t *ip,
2417 struct xfs_defer_ops *dfops)
2418{
2419 int error;
2420 struct xfs_icluster xic = { 0 };
2421
2422 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
2423 ASSERT(VFS_I(ip)->i_nlink == 0);
2424 ASSERT(ip->i_d.di_nextents == 0);
2425 ASSERT(ip->i_d.di_anextents == 0);
2426 ASSERT(ip->i_d.di_size == 0 || !S_ISREG(VFS_I(ip)->i_mode));
2427 ASSERT(ip->i_d.di_nblocks == 0);
2428
2429 /*
2430 * Pull the on-disk inode from the AGI unlinked list.
2431 */
2432 error = xfs_iunlink_remove(tp, ip);
2433 if (error)
2434 return error;
2435
2436 error = xfs_difree(tp, ip->i_ino, dfops, &xic);
2437 if (error)
2438 return error;
2439
2440 VFS_I(ip)->i_mode = 0; /* mark incore inode as free */
2441 ip->i_d.di_flags = 0;
2442 ip->i_d.di_dmevmask = 0;
2443 ip->i_d.di_forkoff = 0; /* mark the attr fork not in use */
2444 ip->i_d.di_format = XFS_DINODE_FMT_EXTENTS;
2445 ip->i_d.di_aformat = XFS_DINODE_FMT_EXTENTS;
2446 /*
2447 * Bump the generation count so no one will be confused
2448 * by reincarnations of this inode.
2449 */
2450 VFS_I(ip)->i_generation++;
2451 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2452
2453 if (xic.deleted)
2454 error = xfs_ifree_cluster(ip, tp, &xic);
2455
2456 return error;
2457}
2458
2459/*
2460 * This is called to unpin an inode. The caller must have the inode locked
2461 * in at least shared mode so that the buffer cannot be subsequently pinned
2462 * once someone is waiting for it to be unpinned.
2463 */
2464static void
2465xfs_iunpin(
2466 struct xfs_inode *ip)
2467{
2468 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
2469
2470 trace_xfs_inode_unpin_nowait(ip, _RET_IP_);
2471
2472 /* Give the log a push to start the unpinning I/O */
2473 xfs_log_force_lsn(ip->i_mount, ip->i_itemp->ili_last_lsn, 0);
2474
2475}
2476
2477static void
2478__xfs_iunpin_wait(
2479 struct xfs_inode *ip)
2480{
2481 wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_IPINNED_BIT);
2482 DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IPINNED_BIT);
2483
2484 xfs_iunpin(ip);
2485
2486 do {
2487 prepare_to_wait(wq, &wait.wait, TASK_UNINTERRUPTIBLE);
2488 if (xfs_ipincount(ip))
2489 io_schedule();
2490 } while (xfs_ipincount(ip));
2491 finish_wait(wq, &wait.wait);
2492}
2493
2494void
2495xfs_iunpin_wait(
2496 struct xfs_inode *ip)
2497{
2498 if (xfs_ipincount(ip))
2499 __xfs_iunpin_wait(ip);
2500}
2501
2502/*
2503 * Removing an inode from the namespace involves removing the directory entry
2504 * and dropping the link count on the inode. Removing the directory entry can
2505 * result in locking an AGF (directory blocks were freed) and removing a link
2506 * count can result in placing the inode on an unlinked list which results in
2507 * locking an AGI.
2508 *
2509 * The big problem here is that we have an ordering constraint on AGF and AGI
2510 * locking - inode allocation locks the AGI, then can allocate a new extent for
2511 * new inodes, locking the AGF after the AGI. Similarly, freeing the inode
2512 * removes the inode from the unlinked list, requiring that we lock the AGI
2513 * first, and then freeing the inode can result in an inode chunk being freed
2514 * and hence freeing disk space requiring that we lock an AGF.
2515 *
2516 * Hence the ordering that is imposed by other parts of the code is AGI before
2517 * AGF. This means we cannot remove the directory entry before we drop the inode
2518 * reference count and put it on the unlinked list as this results in a lock
2519 * order of AGF then AGI, and this can deadlock against inode allocation and
2520 * freeing. Therefore we must drop the link counts before we remove the
2521 * directory entry.
2522 *
2523 * This is still safe from a transactional point of view - it is not until we
2524 * get to xfs_defer_finish() that we have the possibility of multiple
2525 * transactions in this operation. Hence as long as we remove the directory
2526 * entry and drop the link count in the first transaction of the remove
2527 * operation, there are no transactional constraints on the ordering here.
2528 */
2529int
2530xfs_remove(
2531 xfs_inode_t *dp,
2532 struct xfs_name *name,
2533 xfs_inode_t *ip)
2534{
2535 xfs_mount_t *mp = dp->i_mount;
2536 xfs_trans_t *tp = NULL;
2537 int is_dir = S_ISDIR(VFS_I(ip)->i_mode);
2538 int error = 0;
2539 struct xfs_defer_ops dfops;
2540 xfs_fsblock_t first_block;
2541 uint resblks;
2542
2543 trace_xfs_remove(dp, name);
2544
2545 if (XFS_FORCED_SHUTDOWN(mp))
2546 return -EIO;
2547
2548 error = xfs_qm_dqattach(dp, 0);
2549 if (error)
2550 goto std_return;
2551
2552 error = xfs_qm_dqattach(ip, 0);
2553 if (error)
2554 goto std_return;
2555
2556 /*
2557 * We try to get the real space reservation first,
2558 * allowing for directory btree deletion(s) implying
2559 * possible bmap insert(s). If we can't get the space
2560 * reservation then we use 0 instead, and avoid the bmap
2561 * btree insert(s) in the directory code by, if the bmap
2562 * insert tries to happen, instead trimming the LAST
2563 * block from the directory.
2564 */
2565 resblks = XFS_REMOVE_SPACE_RES(mp);
2566 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_remove, resblks, 0, 0, &tp);
2567 if (error == -ENOSPC) {
2568 resblks = 0;
2569 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_remove, 0, 0, 0,
2570 &tp);
2571 }
2572 if (error) {
2573 ASSERT(error != -ENOSPC);
2574 goto std_return;
2575 }
2576
2577 xfs_lock_two_inodes(dp, ip, XFS_ILOCK_EXCL);
2578
2579 xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
2580 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
2581
2582 /*
2583 * If we're removing a directory perform some additional validation.
2584 */
2585 if (is_dir) {
2586 ASSERT(VFS_I(ip)->i_nlink >= 2);
2587 if (VFS_I(ip)->i_nlink != 2) {
2588 error = -ENOTEMPTY;
2589 goto out_trans_cancel;
2590 }
2591 if (!xfs_dir_isempty(ip)) {
2592 error = -ENOTEMPTY;
2593 goto out_trans_cancel;
2594 }
2595
2596 /* Drop the link from ip's "..". */
2597 error = xfs_droplink(tp, dp);
2598 if (error)
2599 goto out_trans_cancel;
2600
2601 /* Drop the "." link from ip to self. */
2602 error = xfs_droplink(tp, ip);
2603 if (error)
2604 goto out_trans_cancel;
2605 } else {
2606 /*
2607 * When removing a non-directory we need to log the parent
2608 * inode here. For a directory this is done implicitly
2609 * by the xfs_droplink call for the ".." entry.
2610 */
2611 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
2612 }
2613 xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2614
2615 /* Drop the link from dp to ip. */
2616 error = xfs_droplink(tp, ip);
2617 if (error)
2618 goto out_trans_cancel;
2619
2620 xfs_defer_init(&dfops, &first_block);
2621 error = xfs_dir_removename(tp, dp, name, ip->i_ino,
2622 &first_block, &dfops, resblks);
2623 if (error) {
2624 ASSERT(error != -ENOENT);
2625 goto out_bmap_cancel;
2626 }
2627
2628 /*
2629 * If this is a synchronous mount, make sure that the
2630 * remove transaction goes to disk before returning to
2631 * the user.
2632 */
2633 if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
2634 xfs_trans_set_sync(tp);
2635
2636 error = xfs_defer_finish(&tp, &dfops, NULL);
2637 if (error)
2638 goto out_bmap_cancel;
2639
2640 error = xfs_trans_commit(tp);
2641 if (error)
2642 goto std_return;
2643
2644 if (is_dir && xfs_inode_is_filestream(ip))
2645 xfs_filestream_deassociate(ip);
2646
2647 return 0;
2648
2649 out_bmap_cancel:
2650 xfs_defer_cancel(&dfops);
2651 out_trans_cancel:
2652 xfs_trans_cancel(tp);
2653 std_return:
2654 return error;
2655}
2656
2657/*
2658 * Enter all inodes for a rename transaction into a sorted array.
2659 */
2660#define __XFS_SORT_INODES 5
2661STATIC void
2662xfs_sort_for_rename(
2663 struct xfs_inode *dp1, /* in: old (source) directory inode */
2664 struct xfs_inode *dp2, /* in: new (target) directory inode */
2665 struct xfs_inode *ip1, /* in: inode of old entry */
2666 struct xfs_inode *ip2, /* in: inode of new entry */
2667 struct xfs_inode *wip, /* in: whiteout inode */
2668 struct xfs_inode **i_tab,/* out: sorted array of inodes */
2669 int *num_inodes) /* in/out: inodes in array */
2670{
2671 int i, j;
2672
2673 ASSERT(*num_inodes == __XFS_SORT_INODES);
2674 memset(i_tab, 0, *num_inodes * sizeof(struct xfs_inode *));
2675
2676 /*
2677 * i_tab contains a list of pointers to inodes. We initialize
2678 * the table here & we'll sort it. We will then use it to
2679 * order the acquisition of the inode locks.
2680 *
2681 * Note that the table may contain duplicates. e.g., dp1 == dp2.
2682 */
2683 i = 0;
2684 i_tab[i++] = dp1;
2685 i_tab[i++] = dp2;
2686 i_tab[i++] = ip1;
2687 if (ip2)
2688 i_tab[i++] = ip2;
2689 if (wip)
2690 i_tab[i++] = wip;
2691 *num_inodes = i;
2692
2693 /*
2694 * Sort the elements via bubble sort. (Remember, there are at
2695 * most 5 elements to sort, so this is adequate.)
2696 */
2697 for (i = 0; i < *num_inodes; i++) {
2698 for (j = 1; j < *num_inodes; j++) {
2699 if (i_tab[j]->i_ino < i_tab[j-1]->i_ino) {
2700 struct xfs_inode *temp = i_tab[j];
2701 i_tab[j] = i_tab[j-1];
2702 i_tab[j-1] = temp;
2703 }
2704 }
2705 }
2706}
2707
2708static int
2709xfs_finish_rename(
2710 struct xfs_trans *tp,
2711 struct xfs_defer_ops *dfops)
2712{
2713 int error;
2714
2715 /*
2716 * If this is a synchronous mount, make sure that the rename transaction
2717 * goes to disk before returning to the user.
2718 */
2719 if (tp->t_mountp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
2720 xfs_trans_set_sync(tp);
2721
2722 error = xfs_defer_finish(&tp, dfops, NULL);
2723 if (error) {
2724 xfs_defer_cancel(dfops);
2725 xfs_trans_cancel(tp);
2726 return error;
2727 }
2728
2729 return xfs_trans_commit(tp);
2730}
2731
2732/*
2733 * xfs_cross_rename()
2734 *
2735 * responsible for handling RENAME_EXCHANGE flag in renameat2() sytemcall
2736 */
2737STATIC int
2738xfs_cross_rename(
2739 struct xfs_trans *tp,
2740 struct xfs_inode *dp1,
2741 struct xfs_name *name1,
2742 struct xfs_inode *ip1,
2743 struct xfs_inode *dp2,
2744 struct xfs_name *name2,
2745 struct xfs_inode *ip2,
2746 struct xfs_defer_ops *dfops,
2747 xfs_fsblock_t *first_block,
2748 int spaceres)
2749{
2750 int error = 0;
2751 int ip1_flags = 0;
2752 int ip2_flags = 0;
2753 int dp2_flags = 0;
2754
2755 /* Swap inode number for dirent in first parent */
2756 error = xfs_dir_replace(tp, dp1, name1,
2757 ip2->i_ino,
2758 first_block, dfops, spaceres);
2759 if (error)
2760 goto out_trans_abort;
2761
2762 /* Swap inode number for dirent in second parent */
2763 error = xfs_dir_replace(tp, dp2, name2,
2764 ip1->i_ino,
2765 first_block, dfops, spaceres);
2766 if (error)
2767 goto out_trans_abort;
2768
2769 /*
2770 * If we're renaming one or more directories across different parents,
2771 * update the respective ".." entries (and link counts) to match the new
2772 * parents.
2773 */
2774 if (dp1 != dp2) {
2775 dp2_flags = XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2776
2777 if (S_ISDIR(VFS_I(ip2)->i_mode)) {
2778 error = xfs_dir_replace(tp, ip2, &xfs_name_dotdot,
2779 dp1->i_ino, first_block,
2780 dfops, spaceres);
2781 if (error)
2782 goto out_trans_abort;
2783
2784 /* transfer ip2 ".." reference to dp1 */
2785 if (!S_ISDIR(VFS_I(ip1)->i_mode)) {
2786 error = xfs_droplink(tp, dp2);
2787 if (error)
2788 goto out_trans_abort;
2789 error = xfs_bumplink(tp, dp1);
2790 if (error)
2791 goto out_trans_abort;
2792 }
2793
2794 /*
2795 * Although ip1 isn't changed here, userspace needs
2796 * to be warned about the change, so that applications
2797 * relying on it (like backup ones), will properly
2798 * notify the change
2799 */
2800 ip1_flags |= XFS_ICHGTIME_CHG;
2801 ip2_flags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2802 }
2803
2804 if (S_ISDIR(VFS_I(ip1)->i_mode)) {
2805 error = xfs_dir_replace(tp, ip1, &xfs_name_dotdot,
2806 dp2->i_ino, first_block,
2807 dfops, spaceres);
2808 if (error)
2809 goto out_trans_abort;
2810
2811 /* transfer ip1 ".." reference to dp2 */
2812 if (!S_ISDIR(VFS_I(ip2)->i_mode)) {
2813 error = xfs_droplink(tp, dp1);
2814 if (error)
2815 goto out_trans_abort;
2816 error = xfs_bumplink(tp, dp2);
2817 if (error)
2818 goto out_trans_abort;
2819 }
2820
2821 /*
2822 * Although ip2 isn't changed here, userspace needs
2823 * to be warned about the change, so that applications
2824 * relying on it (like backup ones), will properly
2825 * notify the change
2826 */
2827 ip1_flags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2828 ip2_flags |= XFS_ICHGTIME_CHG;
2829 }
2830 }
2831
2832 if (ip1_flags) {
2833 xfs_trans_ichgtime(tp, ip1, ip1_flags);
2834 xfs_trans_log_inode(tp, ip1, XFS_ILOG_CORE);
2835 }
2836 if (ip2_flags) {
2837 xfs_trans_ichgtime(tp, ip2, ip2_flags);
2838 xfs_trans_log_inode(tp, ip2, XFS_ILOG_CORE);
2839 }
2840 if (dp2_flags) {
2841 xfs_trans_ichgtime(tp, dp2, dp2_flags);
2842 xfs_trans_log_inode(tp, dp2, XFS_ILOG_CORE);
2843 }
2844 xfs_trans_ichgtime(tp, dp1, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2845 xfs_trans_log_inode(tp, dp1, XFS_ILOG_CORE);
2846 return xfs_finish_rename(tp, dfops);
2847
2848out_trans_abort:
2849 xfs_defer_cancel(dfops);
2850 xfs_trans_cancel(tp);
2851 return error;
2852}
2853
2854/*
2855 * xfs_rename_alloc_whiteout()
2856 *
2857 * Return a referenced, unlinked, unlocked inode that that can be used as a
2858 * whiteout in a rename transaction. We use a tmpfile inode here so that if we
2859 * crash between allocating the inode and linking it into the rename transaction
2860 * recovery will free the inode and we won't leak it.
2861 */
2862static int
2863xfs_rename_alloc_whiteout(
2864 struct xfs_inode *dp,
2865 struct xfs_inode **wip)
2866{
2867 struct xfs_inode *tmpfile;
2868 int error;
2869
2870 error = xfs_create_tmpfile(dp, NULL, S_IFCHR | WHITEOUT_MODE, &tmpfile);
2871 if (error)
2872 return error;
2873
2874 /*
2875 * Prepare the tmpfile inode as if it were created through the VFS.
2876 * Otherwise, the link increment paths will complain about nlink 0->1.
2877 * Drop the link count as done by d_tmpfile(), complete the inode setup
2878 * and flag it as linkable.
2879 */
2880 drop_nlink(VFS_I(tmpfile));
2881 xfs_setup_iops(tmpfile);
2882 xfs_finish_inode_setup(tmpfile);
2883 VFS_I(tmpfile)->i_state |= I_LINKABLE;
2884
2885 *wip = tmpfile;
2886 return 0;
2887}
2888
2889/*
2890 * xfs_rename
2891 */
2892int
2893xfs_rename(
2894 struct xfs_inode *src_dp,
2895 struct xfs_name *src_name,
2896 struct xfs_inode *src_ip,
2897 struct xfs_inode *target_dp,
2898 struct xfs_name *target_name,
2899 struct xfs_inode *target_ip,
2900 unsigned int flags)
2901{
2902 struct xfs_mount *mp = src_dp->i_mount;
2903 struct xfs_trans *tp;
2904 struct xfs_defer_ops dfops;
2905 xfs_fsblock_t first_block;
2906 struct xfs_inode *wip = NULL; /* whiteout inode */
2907 struct xfs_inode *inodes[__XFS_SORT_INODES];
2908 int num_inodes = __XFS_SORT_INODES;
2909 bool new_parent = (src_dp != target_dp);
2910 bool src_is_directory = S_ISDIR(VFS_I(src_ip)->i_mode);
2911 int spaceres;
2912 int error;
2913
2914 trace_xfs_rename(src_dp, target_dp, src_name, target_name);
2915
2916 if ((flags & RENAME_EXCHANGE) && !target_ip)
2917 return -EINVAL;
2918
2919 /*
2920 * If we are doing a whiteout operation, allocate the whiteout inode
2921 * we will be placing at the target and ensure the type is set
2922 * appropriately.
2923 */
2924 if (flags & RENAME_WHITEOUT) {
2925 ASSERT(!(flags & (RENAME_NOREPLACE | RENAME_EXCHANGE)));
2926 error = xfs_rename_alloc_whiteout(target_dp, &wip);
2927 if (error)
2928 return error;
2929
2930 /* setup target dirent info as whiteout */
2931 src_name->type = XFS_DIR3_FT_CHRDEV;
2932 }
2933
2934 xfs_sort_for_rename(src_dp, target_dp, src_ip, target_ip, wip,
2935 inodes, &num_inodes);
2936
2937 spaceres = XFS_RENAME_SPACE_RES(mp, target_name->len);
2938 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, spaceres, 0, 0, &tp);
2939 if (error == -ENOSPC) {
2940 spaceres = 0;
2941 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, 0, 0, 0,
2942 &tp);
2943 }
2944 if (error)
2945 goto out_release_wip;
2946
2947 /*
2948 * Attach the dquots to the inodes
2949 */
2950 error = xfs_qm_vop_rename_dqattach(inodes);
2951 if (error)
2952 goto out_trans_cancel;
2953
2954 /*
2955 * Lock all the participating inodes. Depending upon whether
2956 * the target_name exists in the target directory, and
2957 * whether the target directory is the same as the source
2958 * directory, we can lock from 2 to 4 inodes.
2959 */
2960 xfs_lock_inodes(inodes, num_inodes, XFS_ILOCK_EXCL);
2961
2962 /*
2963 * Join all the inodes to the transaction. From this point on,
2964 * we can rely on either trans_commit or trans_cancel to unlock
2965 * them.
2966 */
2967 xfs_trans_ijoin(tp, src_dp, XFS_ILOCK_EXCL);
2968 if (new_parent)
2969 xfs_trans_ijoin(tp, target_dp, XFS_ILOCK_EXCL);
2970 xfs_trans_ijoin(tp, src_ip, XFS_ILOCK_EXCL);
2971 if (target_ip)
2972 xfs_trans_ijoin(tp, target_ip, XFS_ILOCK_EXCL);
2973 if (wip)
2974 xfs_trans_ijoin(tp, wip, XFS_ILOCK_EXCL);
2975
2976 /*
2977 * If we are using project inheritance, we only allow renames
2978 * into our tree when the project IDs are the same; else the
2979 * tree quota mechanism would be circumvented.
2980 */
2981 if (unlikely((target_dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
2982 (xfs_get_projid(target_dp) != xfs_get_projid(src_ip)))) {
2983 error = -EXDEV;
2984 goto out_trans_cancel;
2985 }
2986
2987 xfs_defer_init(&dfops, &first_block);
2988
2989 /* RENAME_EXCHANGE is unique from here on. */
2990 if (flags & RENAME_EXCHANGE)
2991 return xfs_cross_rename(tp, src_dp, src_name, src_ip,
2992 target_dp, target_name, target_ip,
2993 &dfops, &first_block, spaceres);
2994
2995 /*
2996 * Set up the target.
2997 */
2998 if (target_ip == NULL) {
2999 /*
3000 * If there's no space reservation, check the entry will
3001 * fit before actually inserting it.
3002 */
3003 if (!spaceres) {
3004 error = xfs_dir_canenter(tp, target_dp, target_name);
3005 if (error)
3006 goto out_trans_cancel;
3007 }
3008 /*
3009 * If target does not exist and the rename crosses
3010 * directories, adjust the target directory link count
3011 * to account for the ".." reference from the new entry.
3012 */
3013 error = xfs_dir_createname(tp, target_dp, target_name,
3014 src_ip->i_ino, &first_block,
3015 &dfops, spaceres);
3016 if (error)
3017 goto out_bmap_cancel;
3018
3019 xfs_trans_ichgtime(tp, target_dp,
3020 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3021
3022 if (new_parent && src_is_directory) {
3023 error = xfs_bumplink(tp, target_dp);
3024 if (error)
3025 goto out_bmap_cancel;
3026 }
3027 } else { /* target_ip != NULL */
3028 /*
3029 * If target exists and it's a directory, check that both
3030 * target and source are directories and that target can be
3031 * destroyed, or that neither is a directory.
3032 */
3033 if (S_ISDIR(VFS_I(target_ip)->i_mode)) {
3034 /*
3035 * Make sure target dir is empty.
3036 */
3037 if (!(xfs_dir_isempty(target_ip)) ||
3038 (VFS_I(target_ip)->i_nlink > 2)) {
3039 error = -EEXIST;
3040 goto out_trans_cancel;
3041 }
3042 }
3043
3044 /*
3045 * Link the source inode under the target name.
3046 * If the source inode is a directory and we are moving
3047 * it across directories, its ".." entry will be
3048 * inconsistent until we replace that down below.
3049 *
3050 * In case there is already an entry with the same
3051 * name at the destination directory, remove it first.
3052 */
3053 error = xfs_dir_replace(tp, target_dp, target_name,
3054 src_ip->i_ino,
3055 &first_block, &dfops, spaceres);
3056 if (error)
3057 goto out_bmap_cancel;
3058
3059 xfs_trans_ichgtime(tp, target_dp,
3060 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3061
3062 /*
3063 * Decrement the link count on the target since the target
3064 * dir no longer points to it.
3065 */
3066 error = xfs_droplink(tp, target_ip);
3067 if (error)
3068 goto out_bmap_cancel;
3069
3070 if (src_is_directory) {
3071 /*
3072 * Drop the link from the old "." entry.
3073 */
3074 error = xfs_droplink(tp, target_ip);
3075 if (error)
3076 goto out_bmap_cancel;
3077 }
3078 } /* target_ip != NULL */
3079
3080 /*
3081 * Remove the source.
3082 */
3083 if (new_parent && src_is_directory) {
3084 /*
3085 * Rewrite the ".." entry to point to the new
3086 * directory.
3087 */
3088 error = xfs_dir_replace(tp, src_ip, &xfs_name_dotdot,
3089 target_dp->i_ino,
3090 &first_block, &dfops, spaceres);
3091 ASSERT(error != -EEXIST);
3092 if (error)
3093 goto out_bmap_cancel;
3094 }
3095
3096 /*
3097 * We always want to hit the ctime on the source inode.
3098 *
3099 * This isn't strictly required by the standards since the source
3100 * inode isn't really being changed, but old unix file systems did
3101 * it and some incremental backup programs won't work without it.
3102 */
3103 xfs_trans_ichgtime(tp, src_ip, XFS_ICHGTIME_CHG);
3104 xfs_trans_log_inode(tp, src_ip, XFS_ILOG_CORE);
3105
3106 /*
3107 * Adjust the link count on src_dp. This is necessary when
3108 * renaming a directory, either within one parent when
3109 * the target existed, or across two parent directories.
3110 */
3111 if (src_is_directory && (new_parent || target_ip != NULL)) {
3112
3113 /*
3114 * Decrement link count on src_directory since the
3115 * entry that's moved no longer points to it.
3116 */
3117 error = xfs_droplink(tp, src_dp);
3118 if (error)
3119 goto out_bmap_cancel;
3120 }
3121
3122 /*
3123 * For whiteouts, we only need to update the source dirent with the
3124 * inode number of the whiteout inode rather than removing it
3125 * altogether.
3126 */
3127 if (wip) {
3128 error = xfs_dir_replace(tp, src_dp, src_name, wip->i_ino,
3129 &first_block, &dfops, spaceres);
3130 } else
3131 error = xfs_dir_removename(tp, src_dp, src_name, src_ip->i_ino,
3132 &first_block, &dfops, spaceres);
3133 if (error)
3134 goto out_bmap_cancel;
3135
3136 /*
3137 * For whiteouts, we need to bump the link count on the whiteout inode.
3138 * This means that failures all the way up to this point leave the inode
3139 * on the unlinked list and so cleanup is a simple matter of dropping
3140 * the remaining reference to it. If we fail here after bumping the link
3141 * count, we're shutting down the filesystem so we'll never see the
3142 * intermediate state on disk.
3143 */
3144 if (wip) {
3145 ASSERT(VFS_I(wip)->i_nlink == 0);
3146 error = xfs_bumplink(tp, wip);
3147 if (error)
3148 goto out_bmap_cancel;
3149 error = xfs_iunlink_remove(tp, wip);
3150 if (error)
3151 goto out_bmap_cancel;
3152 xfs_trans_log_inode(tp, wip, XFS_ILOG_CORE);
3153
3154 /*
3155 * Now we have a real link, clear the "I'm a tmpfile" state
3156 * flag from the inode so it doesn't accidentally get misused in
3157 * future.
3158 */
3159 VFS_I(wip)->i_state &= ~I_LINKABLE;
3160 }
3161
3162 xfs_trans_ichgtime(tp, src_dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3163 xfs_trans_log_inode(tp, src_dp, XFS_ILOG_CORE);
3164 if (new_parent)
3165 xfs_trans_log_inode(tp, target_dp, XFS_ILOG_CORE);
3166
3167 error = xfs_finish_rename(tp, &dfops);
3168 if (wip)
3169 IRELE(wip);
3170 return error;
3171
3172out_bmap_cancel:
3173 xfs_defer_cancel(&dfops);
3174out_trans_cancel:
3175 xfs_trans_cancel(tp);
3176out_release_wip:
3177 if (wip)
3178 IRELE(wip);
3179 return error;
3180}
3181
3182STATIC int
3183xfs_iflush_cluster(
3184 struct xfs_inode *ip,
3185 struct xfs_buf *bp)
3186{
3187 struct xfs_mount *mp = ip->i_mount;
3188 struct xfs_perag *pag;
3189 unsigned long first_index, mask;
3190 unsigned long inodes_per_cluster;
3191 int cilist_size;
3192 struct xfs_inode **cilist;
3193 struct xfs_inode *cip;
3194 int nr_found;
3195 int clcount = 0;
3196 int bufwasdelwri;
3197 int i;
3198
3199 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
3200
3201 inodes_per_cluster = mp->m_inode_cluster_size >> mp->m_sb.sb_inodelog;
3202 cilist_size = inodes_per_cluster * sizeof(xfs_inode_t *);
3203 cilist = kmem_alloc(cilist_size, KM_MAYFAIL|KM_NOFS);
3204 if (!cilist)
3205 goto out_put;
3206
3207 mask = ~(((mp->m_inode_cluster_size >> mp->m_sb.sb_inodelog)) - 1);
3208 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino) & mask;
3209 rcu_read_lock();
3210 /* really need a gang lookup range call here */
3211 nr_found = radix_tree_gang_lookup(&pag->pag_ici_root, (void**)cilist,
3212 first_index, inodes_per_cluster);
3213 if (nr_found == 0)
3214 goto out_free;
3215
3216 for (i = 0; i < nr_found; i++) {
3217 cip = cilist[i];
3218 if (cip == ip)
3219 continue;
3220
3221 /*
3222 * because this is an RCU protected lookup, we could find a
3223 * recently freed or even reallocated inode during the lookup.
3224 * We need to check under the i_flags_lock for a valid inode
3225 * here. Skip it if it is not valid or the wrong inode.
3226 */
3227 spin_lock(&cip->i_flags_lock);
3228 if (!cip->i_ino ||
3229 __xfs_iflags_test(cip, XFS_ISTALE)) {
3230 spin_unlock(&cip->i_flags_lock);
3231 continue;
3232 }
3233
3234 /*
3235 * Once we fall off the end of the cluster, no point checking
3236 * any more inodes in the list because they will also all be
3237 * outside the cluster.
3238 */
3239 if ((XFS_INO_TO_AGINO(mp, cip->i_ino) & mask) != first_index) {
3240 spin_unlock(&cip->i_flags_lock);
3241 break;
3242 }
3243 spin_unlock(&cip->i_flags_lock);
3244
3245 /*
3246 * Do an un-protected check to see if the inode is dirty and
3247 * is a candidate for flushing. These checks will be repeated
3248 * later after the appropriate locks are acquired.
3249 */
3250 if (xfs_inode_clean(cip) && xfs_ipincount(cip) == 0)
3251 continue;
3252
3253 /*
3254 * Try to get locks. If any are unavailable or it is pinned,
3255 * then this inode cannot be flushed and is skipped.
3256 */
3257
3258 if (!xfs_ilock_nowait(cip, XFS_ILOCK_SHARED))
3259 continue;
3260 if (!xfs_iflock_nowait(cip)) {
3261 xfs_iunlock(cip, XFS_ILOCK_SHARED);
3262 continue;
3263 }
3264 if (xfs_ipincount(cip)) {
3265 xfs_ifunlock(cip);
3266 xfs_iunlock(cip, XFS_ILOCK_SHARED);
3267 continue;
3268 }
3269
3270
3271 /*
3272 * Check the inode number again, just to be certain we are not
3273 * racing with freeing in xfs_reclaim_inode(). See the comments
3274 * in that function for more information as to why the initial
3275 * check is not sufficient.
3276 */
3277 if (!cip->i_ino) {
3278 xfs_ifunlock(cip);
3279 xfs_iunlock(cip, XFS_ILOCK_SHARED);
3280 continue;
3281 }
3282
3283 /*
3284 * arriving here means that this inode can be flushed. First
3285 * re-check that it's dirty before flushing.
3286 */
3287 if (!xfs_inode_clean(cip)) {
3288 int error;
3289 error = xfs_iflush_int(cip, bp);
3290 if (error) {
3291 xfs_iunlock(cip, XFS_ILOCK_SHARED);
3292 goto cluster_corrupt_out;
3293 }
3294 clcount++;
3295 } else {
3296 xfs_ifunlock(cip);
3297 }
3298 xfs_iunlock(cip, XFS_ILOCK_SHARED);
3299 }
3300
3301 if (clcount) {
3302 XFS_STATS_INC(mp, xs_icluster_flushcnt);
3303 XFS_STATS_ADD(mp, xs_icluster_flushinode, clcount);
3304 }
3305
3306out_free:
3307 rcu_read_unlock();
3308 kmem_free(cilist);
3309out_put:
3310 xfs_perag_put(pag);
3311 return 0;
3312
3313
3314cluster_corrupt_out:
3315 /*
3316 * Corruption detected in the clustering loop. Invalidate the
3317 * inode buffer and shut down the filesystem.
3318 */
3319 rcu_read_unlock();
3320 /*
3321 * Clean up the buffer. If it was delwri, just release it --
3322 * brelse can handle it with no problems. If not, shut down the
3323 * filesystem before releasing the buffer.
3324 */
3325 bufwasdelwri = (bp->b_flags & _XBF_DELWRI_Q);
3326 if (bufwasdelwri)
3327 xfs_buf_relse(bp);
3328
3329 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
3330
3331 if (!bufwasdelwri) {
3332 /*
3333 * Just like incore_relse: if we have b_iodone functions,
3334 * mark the buffer as an error and call them. Otherwise
3335 * mark it as stale and brelse.
3336 */
3337 if (bp->b_iodone) {
3338 bp->b_flags &= ~XBF_DONE;
3339 xfs_buf_stale(bp);
3340 xfs_buf_ioerror(bp, -EIO);
3341 xfs_buf_ioend(bp);
3342 } else {
3343 xfs_buf_stale(bp);
3344 xfs_buf_relse(bp);
3345 }
3346 }
3347
3348 /*
3349 * Unlocks the flush lock
3350 */
3351 xfs_iflush_abort(cip, false);
3352 kmem_free(cilist);
3353 xfs_perag_put(pag);
3354 return -EFSCORRUPTED;
3355}
3356
3357/*
3358 * Flush dirty inode metadata into the backing buffer.
3359 *
3360 * The caller must have the inode lock and the inode flush lock held. The
3361 * inode lock will still be held upon return to the caller, and the inode
3362 * flush lock will be released after the inode has reached the disk.
3363 *
3364 * The caller must write out the buffer returned in *bpp and release it.
3365 */
3366int
3367xfs_iflush(
3368 struct xfs_inode *ip,
3369 struct xfs_buf **bpp)
3370{
3371 struct xfs_mount *mp = ip->i_mount;
3372 struct xfs_buf *bp = NULL;
3373 struct xfs_dinode *dip;
3374 int error;
3375
3376 XFS_STATS_INC(mp, xs_iflush_count);
3377
3378 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
3379 ASSERT(xfs_isiflocked(ip));
3380 ASSERT(ip->i_d.di_format != XFS_DINODE_FMT_BTREE ||
3381 ip->i_d.di_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
3382
3383 *bpp = NULL;
3384
3385 xfs_iunpin_wait(ip);
3386
3387 /*
3388 * For stale inodes we cannot rely on the backing buffer remaining
3389 * stale in cache for the remaining life of the stale inode and so
3390 * xfs_imap_to_bp() below may give us a buffer that no longer contains
3391 * inodes below. We have to check this after ensuring the inode is
3392 * unpinned so that it is safe to reclaim the stale inode after the
3393 * flush call.
3394 */
3395 if (xfs_iflags_test(ip, XFS_ISTALE)) {
3396 xfs_ifunlock(ip);
3397 return 0;
3398 }
3399
3400 /*
3401 * This may have been unpinned because the filesystem is shutting
3402 * down forcibly. If that's the case we must not write this inode
3403 * to disk, because the log record didn't make it to disk.
3404 *
3405 * We also have to remove the log item from the AIL in this case,
3406 * as we wait for an empty AIL as part of the unmount process.
3407 */
3408 if (XFS_FORCED_SHUTDOWN(mp)) {
3409 error = -EIO;
3410 goto abort_out;
3411 }
3412
3413 /*
3414 * Get the buffer containing the on-disk inode. We are doing a try-lock
3415 * operation here, so we may get an EAGAIN error. In that case, we
3416 * simply want to return with the inode still dirty.
3417 *
3418 * If we get any other error, we effectively have a corruption situation
3419 * and we cannot flush the inode, so we treat it the same as failing
3420 * xfs_iflush_int().
3421 */
3422 error = xfs_imap_to_bp(mp, NULL, &ip->i_imap, &dip, &bp, XBF_TRYLOCK,
3423 0);
3424 if (error == -EAGAIN) {
3425 xfs_ifunlock(ip);
3426 return error;
3427 }
3428 if (error)
3429 goto corrupt_out;
3430
3431 /*
3432 * First flush out the inode that xfs_iflush was called with.
3433 */
3434 error = xfs_iflush_int(ip, bp);
3435 if (error)
3436 goto corrupt_out;
3437
3438 /*
3439 * If the buffer is pinned then push on the log now so we won't
3440 * get stuck waiting in the write for too long.
3441 */
3442 if (xfs_buf_ispinned(bp))
3443 xfs_log_force(mp, 0);
3444
3445 /*
3446 * inode clustering:
3447 * see if other inodes can be gathered into this write
3448 */
3449 error = xfs_iflush_cluster(ip, bp);
3450 if (error)
3451 goto cluster_corrupt_out;
3452
3453 *bpp = bp;
3454 return 0;
3455
3456corrupt_out:
3457 if (bp)
3458 xfs_buf_relse(bp);
3459 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
3460cluster_corrupt_out:
3461 error = -EFSCORRUPTED;
3462abort_out:
3463 /*
3464 * Unlocks the flush lock
3465 */
3466 xfs_iflush_abort(ip, false);
3467 return error;
3468}
3469
3470STATIC int
3471xfs_iflush_int(
3472 struct xfs_inode *ip,
3473 struct xfs_buf *bp)
3474{
3475 struct xfs_inode_log_item *iip = ip->i_itemp;
3476 struct xfs_dinode *dip;
3477 struct xfs_mount *mp = ip->i_mount;
3478
3479 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
3480 ASSERT(xfs_isiflocked(ip));
3481 ASSERT(ip->i_d.di_format != XFS_DINODE_FMT_BTREE ||
3482 ip->i_d.di_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
3483 ASSERT(iip != NULL && iip->ili_fields != 0);
3484 ASSERT(ip->i_d.di_version > 1);
3485
3486 /* set *dip = inode's place in the buffer */
3487 dip = xfs_buf_offset(bp, ip->i_imap.im_boffset);
3488
3489 if (XFS_TEST_ERROR(dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC),
3490 mp, XFS_ERRTAG_IFLUSH_1, XFS_RANDOM_IFLUSH_1)) {
3491 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3492 "%s: Bad inode %Lu magic number 0x%x, ptr 0x%p",
3493 __func__, ip->i_ino, be16_to_cpu(dip->di_magic), dip);
3494 goto corrupt_out;
3495 }
3496 if (S_ISREG(VFS_I(ip)->i_mode)) {
3497 if (XFS_TEST_ERROR(
3498 (ip->i_d.di_format != XFS_DINODE_FMT_EXTENTS) &&
3499 (ip->i_d.di_format != XFS_DINODE_FMT_BTREE),
3500 mp, XFS_ERRTAG_IFLUSH_3, XFS_RANDOM_IFLUSH_3)) {
3501 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3502 "%s: Bad regular inode %Lu, ptr 0x%p",
3503 __func__, ip->i_ino, ip);
3504 goto corrupt_out;
3505 }
3506 } else if (S_ISDIR(VFS_I(ip)->i_mode)) {
3507 if (XFS_TEST_ERROR(
3508 (ip->i_d.di_format != XFS_DINODE_FMT_EXTENTS) &&
3509 (ip->i_d.di_format != XFS_DINODE_FMT_BTREE) &&
3510 (ip->i_d.di_format != XFS_DINODE_FMT_LOCAL),
3511 mp, XFS_ERRTAG_IFLUSH_4, XFS_RANDOM_IFLUSH_4)) {
3512 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3513 "%s: Bad directory inode %Lu, ptr 0x%p",
3514 __func__, ip->i_ino, ip);
3515 goto corrupt_out;
3516 }
3517 }
3518 if (XFS_TEST_ERROR(ip->i_d.di_nextents + ip->i_d.di_anextents >
3519 ip->i_d.di_nblocks, mp, XFS_ERRTAG_IFLUSH_5,
3520 XFS_RANDOM_IFLUSH_5)) {
3521 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3522 "%s: detected corrupt incore inode %Lu, "
3523 "total extents = %d, nblocks = %Ld, ptr 0x%p",
3524 __func__, ip->i_ino,
3525 ip->i_d.di_nextents + ip->i_d.di_anextents,
3526 ip->i_d.di_nblocks, ip);
3527 goto corrupt_out;
3528 }
3529 if (XFS_TEST_ERROR(ip->i_d.di_forkoff > mp->m_sb.sb_inodesize,
3530 mp, XFS_ERRTAG_IFLUSH_6, XFS_RANDOM_IFLUSH_6)) {
3531 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3532 "%s: bad inode %Lu, forkoff 0x%x, ptr 0x%p",
3533 __func__, ip->i_ino, ip->i_d.di_forkoff, ip);
3534 goto corrupt_out;
3535 }
3536
3537 /*
3538 * Inode item log recovery for v2 inodes are dependent on the
3539 * di_flushiter count for correct sequencing. We bump the flush
3540 * iteration count so we can detect flushes which postdate a log record
3541 * during recovery. This is redundant as we now log every change and
3542 * hence this can't happen but we need to still do it to ensure
3543 * backwards compatibility with old kernels that predate logging all
3544 * inode changes.
3545 */
3546 if (ip->i_d.di_version < 3)
3547 ip->i_d.di_flushiter++;
3548
3549 /*
3550 * Copy the dirty parts of the inode into the on-disk inode. We always
3551 * copy out the core of the inode, because if the inode is dirty at all
3552 * the core must be.
3553 */
3554 xfs_inode_to_disk(ip, dip, iip->ili_item.li_lsn);
3555
3556 /* Wrap, we never let the log put out DI_MAX_FLUSH */
3557 if (ip->i_d.di_flushiter == DI_MAX_FLUSH)
3558 ip->i_d.di_flushiter = 0;
3559
3560 xfs_iflush_fork(ip, dip, iip, XFS_DATA_FORK);
3561 if (XFS_IFORK_Q(ip))
3562 xfs_iflush_fork(ip, dip, iip, XFS_ATTR_FORK);
3563 xfs_inobp_check(mp, bp);
3564
3565 /*
3566 * We've recorded everything logged in the inode, so we'd like to clear
3567 * the ili_fields bits so we don't log and flush things unnecessarily.
3568 * However, we can't stop logging all this information until the data
3569 * we've copied into the disk buffer is written to disk. If we did we
3570 * might overwrite the copy of the inode in the log with all the data
3571 * after re-logging only part of it, and in the face of a crash we
3572 * wouldn't have all the data we need to recover.
3573 *
3574 * What we do is move the bits to the ili_last_fields field. When
3575 * logging the inode, these bits are moved back to the ili_fields field.
3576 * In the xfs_iflush_done() routine we clear ili_last_fields, since we
3577 * know that the information those bits represent is permanently on
3578 * disk. As long as the flush completes before the inode is logged
3579 * again, then both ili_fields and ili_last_fields will be cleared.
3580 *
3581 * We can play with the ili_fields bits here, because the inode lock
3582 * must be held exclusively in order to set bits there and the flush
3583 * lock protects the ili_last_fields bits. Set ili_logged so the flush
3584 * done routine can tell whether or not to look in the AIL. Also, store
3585 * the current LSN of the inode so that we can tell whether the item has
3586 * moved in the AIL from xfs_iflush_done(). In order to read the lsn we
3587 * need the AIL lock, because it is a 64 bit value that cannot be read
3588 * atomically.
3589 */
3590 iip->ili_last_fields = iip->ili_fields;
3591 iip->ili_fields = 0;
3592 iip->ili_fsync_fields = 0;
3593 iip->ili_logged = 1;
3594
3595 xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
3596 &iip->ili_item.li_lsn);
3597
3598 /*
3599 * Attach the function xfs_iflush_done to the inode's
3600 * buffer. This will remove the inode from the AIL
3601 * and unlock the inode's flush lock when the inode is
3602 * completely written to disk.
3603 */
3604 xfs_buf_attach_iodone(bp, xfs_iflush_done, &iip->ili_item);
3605
3606 /* generate the checksum. */
3607 xfs_dinode_calc_crc(mp, dip);
3608
3609 ASSERT(bp->b_fspriv != NULL);
3610 ASSERT(bp->b_iodone != NULL);
3611 return 0;
3612
3613corrupt_out:
3614 return -EFSCORRUPTED;
3615}
1/*
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}