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_inum.h"
27#include "xfs_sb.h"
28#include "xfs_ag.h"
29#include "xfs_mount.h"
30#include "xfs_inode.h"
31#include "xfs_da_format.h"
32#include "xfs_da_btree.h"
33#include "xfs_dir2.h"
34#include "xfs_attr_sf.h"
35#include "xfs_attr.h"
36#include "xfs_trans_space.h"
37#include "xfs_trans.h"
38#include "xfs_buf_item.h"
39#include "xfs_inode_item.h"
40#include "xfs_ialloc.h"
41#include "xfs_bmap.h"
42#include "xfs_bmap_util.h"
43#include "xfs_error.h"
44#include "xfs_quota.h"
45#include "xfs_filestream.h"
46#include "xfs_cksum.h"
47#include "xfs_trace.h"
48#include "xfs_icache.h"
49#include "xfs_symlink.h"
50#include "xfs_trans_priv.h"
51#include "xfs_log.h"
52#include "xfs_bmap_btree.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(xfs_inode_t *, xfs_buf_t *);
63
64STATIC int xfs_iunlink_remove(xfs_trans_t *, xfs_inode_t *);
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 * These two are wrapper routines around the xfs_ilock() routine used to
82 * centralize some grungy code. They are used in places that wish to lock the
83 * inode solely for reading the extents. The reason these places can't just
84 * call xfs_ilock(ip, XFS_ILOCK_SHARED) is that the inode lock also guards to
85 * bringing in of the extents from disk for a file in b-tree format. If the
86 * inode is in b-tree format, then we need to lock the inode exclusively until
87 * the extents are read in. Locking it exclusively all the time would limit
88 * our parallelism unnecessarily, though. What we do instead is check to see
89 * if the extents have been read in yet, and only lock the inode exclusively
90 * if they have not.
91 *
92 * The functions return a value which should be given to the corresponding
93 * xfs_iunlock() call.
94 */
95uint
96xfs_ilock_data_map_shared(
97 struct xfs_inode *ip)
98{
99 uint lock_mode = XFS_ILOCK_SHARED;
100
101 if (ip->i_d.di_format == XFS_DINODE_FMT_BTREE &&
102 (ip->i_df.if_flags & XFS_IFEXTENTS) == 0)
103 lock_mode = XFS_ILOCK_EXCL;
104 xfs_ilock(ip, lock_mode);
105 return lock_mode;
106}
107
108uint
109xfs_ilock_attr_map_shared(
110 struct xfs_inode *ip)
111{
112 uint lock_mode = XFS_ILOCK_SHARED;
113
114 if (ip->i_d.di_aformat == XFS_DINODE_FMT_BTREE &&
115 (ip->i_afp->if_flags & XFS_IFEXTENTS) == 0)
116 lock_mode = XFS_ILOCK_EXCL;
117 xfs_ilock(ip, lock_mode);
118 return lock_mode;
119}
120
121/*
122 * The xfs inode contains 2 locks: a multi-reader lock called the
123 * i_iolock and a multi-reader lock called the i_lock. This routine
124 * allows either or both of the locks to be obtained.
125 *
126 * The 2 locks should always be ordered so that the IO lock is
127 * obtained first in order to prevent deadlock.
128 *
129 * ip -- the inode being locked
130 * lock_flags -- this parameter indicates the inode's locks
131 * to be locked. It can be:
132 * XFS_IOLOCK_SHARED,
133 * XFS_IOLOCK_EXCL,
134 * XFS_ILOCK_SHARED,
135 * XFS_ILOCK_EXCL,
136 * XFS_IOLOCK_SHARED | XFS_ILOCK_SHARED,
137 * XFS_IOLOCK_SHARED | XFS_ILOCK_EXCL,
138 * XFS_IOLOCK_EXCL | XFS_ILOCK_SHARED,
139 * XFS_IOLOCK_EXCL | XFS_ILOCK_EXCL
140 */
141void
142xfs_ilock(
143 xfs_inode_t *ip,
144 uint lock_flags)
145{
146 trace_xfs_ilock(ip, lock_flags, _RET_IP_);
147
148 /*
149 * You can't set both SHARED and EXCL for the same lock,
150 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
151 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
152 */
153 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
154 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
155 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
156 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
157 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_DEP_MASK)) == 0);
158
159 if (lock_flags & XFS_IOLOCK_EXCL)
160 mrupdate_nested(&ip->i_iolock, XFS_IOLOCK_DEP(lock_flags));
161 else if (lock_flags & XFS_IOLOCK_SHARED)
162 mraccess_nested(&ip->i_iolock, XFS_IOLOCK_DEP(lock_flags));
163
164 if (lock_flags & XFS_ILOCK_EXCL)
165 mrupdate_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
166 else if (lock_flags & XFS_ILOCK_SHARED)
167 mraccess_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
168}
169
170/*
171 * This is just like xfs_ilock(), except that the caller
172 * is guaranteed not to sleep. It returns 1 if it gets
173 * the requested locks and 0 otherwise. If the IO lock is
174 * obtained but the inode lock cannot be, then the IO lock
175 * is dropped before returning.
176 *
177 * ip -- the inode being locked
178 * lock_flags -- this parameter indicates the inode's locks to be
179 * to be locked. See the comment for xfs_ilock() for a list
180 * of valid values.
181 */
182int
183xfs_ilock_nowait(
184 xfs_inode_t *ip,
185 uint lock_flags)
186{
187 trace_xfs_ilock_nowait(ip, lock_flags, _RET_IP_);
188
189 /*
190 * You can't set both SHARED and EXCL for the same lock,
191 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
192 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
193 */
194 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
195 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
196 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
197 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
198 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_DEP_MASK)) == 0);
199
200 if (lock_flags & XFS_IOLOCK_EXCL) {
201 if (!mrtryupdate(&ip->i_iolock))
202 goto out;
203 } else if (lock_flags & XFS_IOLOCK_SHARED) {
204 if (!mrtryaccess(&ip->i_iolock))
205 goto out;
206 }
207 if (lock_flags & XFS_ILOCK_EXCL) {
208 if (!mrtryupdate(&ip->i_lock))
209 goto out_undo_iolock;
210 } else if (lock_flags & XFS_ILOCK_SHARED) {
211 if (!mrtryaccess(&ip->i_lock))
212 goto out_undo_iolock;
213 }
214 return 1;
215
216 out_undo_iolock:
217 if (lock_flags & XFS_IOLOCK_EXCL)
218 mrunlock_excl(&ip->i_iolock);
219 else if (lock_flags & XFS_IOLOCK_SHARED)
220 mrunlock_shared(&ip->i_iolock);
221 out:
222 return 0;
223}
224
225/*
226 * xfs_iunlock() is used to drop the inode locks acquired with
227 * xfs_ilock() and xfs_ilock_nowait(). The caller must pass
228 * in the flags given to xfs_ilock() or xfs_ilock_nowait() so
229 * that we know which locks to drop.
230 *
231 * ip -- the inode being unlocked
232 * lock_flags -- this parameter indicates the inode's locks to be
233 * to be unlocked. See the comment for xfs_ilock() for a list
234 * of valid values for this parameter.
235 *
236 */
237void
238xfs_iunlock(
239 xfs_inode_t *ip,
240 uint lock_flags)
241{
242 /*
243 * You can't set both SHARED and EXCL for the same lock,
244 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_ILOCK_SHARED,
245 * and XFS_ILOCK_EXCL are valid values to set in lock_flags.
246 */
247 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
248 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
249 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
250 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
251 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_DEP_MASK)) == 0);
252 ASSERT(lock_flags != 0);
253
254 if (lock_flags & XFS_IOLOCK_EXCL)
255 mrunlock_excl(&ip->i_iolock);
256 else if (lock_flags & XFS_IOLOCK_SHARED)
257 mrunlock_shared(&ip->i_iolock);
258
259 if (lock_flags & XFS_ILOCK_EXCL)
260 mrunlock_excl(&ip->i_lock);
261 else if (lock_flags & XFS_ILOCK_SHARED)
262 mrunlock_shared(&ip->i_lock);
263
264 trace_xfs_iunlock(ip, lock_flags, _RET_IP_);
265}
266
267/*
268 * give up write locks. the i/o lock cannot be held nested
269 * if it is being demoted.
270 */
271void
272xfs_ilock_demote(
273 xfs_inode_t *ip,
274 uint lock_flags)
275{
276 ASSERT(lock_flags & (XFS_IOLOCK_EXCL|XFS_ILOCK_EXCL));
277 ASSERT((lock_flags & ~(XFS_IOLOCK_EXCL|XFS_ILOCK_EXCL)) == 0);
278
279 if (lock_flags & XFS_ILOCK_EXCL)
280 mrdemote(&ip->i_lock);
281 if (lock_flags & XFS_IOLOCK_EXCL)
282 mrdemote(&ip->i_iolock);
283
284 trace_xfs_ilock_demote(ip, lock_flags, _RET_IP_);
285}
286
287#if defined(DEBUG) || defined(XFS_WARN)
288int
289xfs_isilocked(
290 xfs_inode_t *ip,
291 uint lock_flags)
292{
293 if (lock_flags & (XFS_ILOCK_EXCL|XFS_ILOCK_SHARED)) {
294 if (!(lock_flags & XFS_ILOCK_SHARED))
295 return !!ip->i_lock.mr_writer;
296 return rwsem_is_locked(&ip->i_lock.mr_lock);
297 }
298
299 if (lock_flags & (XFS_IOLOCK_EXCL|XFS_IOLOCK_SHARED)) {
300 if (!(lock_flags & XFS_IOLOCK_SHARED))
301 return !!ip->i_iolock.mr_writer;
302 return rwsem_is_locked(&ip->i_iolock.mr_lock);
303 }
304
305 ASSERT(0);
306 return 0;
307}
308#endif
309
310#ifdef DEBUG
311int xfs_locked_n;
312int xfs_small_retries;
313int xfs_middle_retries;
314int xfs_lots_retries;
315int xfs_lock_delays;
316#endif
317
318/*
319 * Bump the subclass so xfs_lock_inodes() acquires each lock with
320 * a different value
321 */
322static inline int
323xfs_lock_inumorder(int lock_mode, int subclass)
324{
325 if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL))
326 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_IOLOCK_SHIFT;
327 if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL))
328 lock_mode |= (subclass + XFS_LOCK_INUMORDER) << XFS_ILOCK_SHIFT;
329
330 return lock_mode;
331}
332
333/*
334 * The following routine will lock n inodes in exclusive mode.
335 * We assume the caller calls us with the inodes in i_ino order.
336 *
337 * We need to detect deadlock where an inode that we lock
338 * is in the AIL and we start waiting for another inode that is locked
339 * by a thread in a long running transaction (such as truncate). This can
340 * result in deadlock since the long running trans might need to wait
341 * for the inode we just locked in order to push the tail and free space
342 * in the log.
343 */
344void
345xfs_lock_inodes(
346 xfs_inode_t **ips,
347 int inodes,
348 uint lock_mode)
349{
350 int attempts = 0, i, j, try_lock;
351 xfs_log_item_t *lp;
352
353 ASSERT(ips && (inodes >= 2)); /* we need at least two */
354
355 try_lock = 0;
356 i = 0;
357
358again:
359 for (; i < inodes; i++) {
360 ASSERT(ips[i]);
361
362 if (i && (ips[i] == ips[i-1])) /* Already locked */
363 continue;
364
365 /*
366 * If try_lock is not set yet, make sure all locked inodes
367 * are not in the AIL.
368 * If any are, set try_lock to be used later.
369 */
370
371 if (!try_lock) {
372 for (j = (i - 1); j >= 0 && !try_lock; j--) {
373 lp = (xfs_log_item_t *)ips[j]->i_itemp;
374 if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
375 try_lock++;
376 }
377 }
378 }
379
380 /*
381 * If any of the previous locks we have locked is in the AIL,
382 * we must TRY to get the second and subsequent locks. If
383 * we can't get any, we must release all we have
384 * and try again.
385 */
386
387 if (try_lock) {
388 /* try_lock must be 0 if i is 0. */
389 /*
390 * try_lock means we have an inode locked
391 * that is in the AIL.
392 */
393 ASSERT(i != 0);
394 if (!xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i))) {
395 attempts++;
396
397 /*
398 * Unlock all previous guys and try again.
399 * xfs_iunlock will try to push the tail
400 * if the inode is in the AIL.
401 */
402
403 for(j = i - 1; j >= 0; j--) {
404
405 /*
406 * Check to see if we've already
407 * unlocked this one.
408 * Not the first one going back,
409 * and the inode ptr is the same.
410 */
411 if ((j != (i - 1)) && ips[j] ==
412 ips[j+1])
413 continue;
414
415 xfs_iunlock(ips[j], lock_mode);
416 }
417
418 if ((attempts % 5) == 0) {
419 delay(1); /* Don't just spin the CPU */
420#ifdef DEBUG
421 xfs_lock_delays++;
422#endif
423 }
424 i = 0;
425 try_lock = 0;
426 goto again;
427 }
428 } else {
429 xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
430 }
431 }
432
433#ifdef DEBUG
434 if (attempts) {
435 if (attempts < 5) xfs_small_retries++;
436 else if (attempts < 100) xfs_middle_retries++;
437 else xfs_lots_retries++;
438 } else {
439 xfs_locked_n++;
440 }
441#endif
442}
443
444/*
445 * xfs_lock_two_inodes() can only be used to lock one type of lock
446 * at a time - the iolock or the ilock, but not both at once. If
447 * we lock both at once, lockdep will report false positives saying
448 * we have violated locking orders.
449 */
450void
451xfs_lock_two_inodes(
452 xfs_inode_t *ip0,
453 xfs_inode_t *ip1,
454 uint lock_mode)
455{
456 xfs_inode_t *temp;
457 int attempts = 0;
458 xfs_log_item_t *lp;
459
460 if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL))
461 ASSERT((lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) == 0);
462 ASSERT(ip0->i_ino != ip1->i_ino);
463
464 if (ip0->i_ino > ip1->i_ino) {
465 temp = ip0;
466 ip0 = ip1;
467 ip1 = temp;
468 }
469
470 again:
471 xfs_ilock(ip0, xfs_lock_inumorder(lock_mode, 0));
472
473 /*
474 * If the first lock we have locked is in the AIL, we must TRY to get
475 * the second lock. If we can't get it, we must release the first one
476 * and try again.
477 */
478 lp = (xfs_log_item_t *)ip0->i_itemp;
479 if (lp && (lp->li_flags & XFS_LI_IN_AIL)) {
480 if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(lock_mode, 1))) {
481 xfs_iunlock(ip0, lock_mode);
482 if ((++attempts % 5) == 0)
483 delay(1); /* Don't just spin the CPU */
484 goto again;
485 }
486 } else {
487 xfs_ilock(ip1, xfs_lock_inumorder(lock_mode, 1));
488 }
489}
490
491
492void
493__xfs_iflock(
494 struct xfs_inode *ip)
495{
496 wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_IFLOCK_BIT);
497 DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IFLOCK_BIT);
498
499 do {
500 prepare_to_wait_exclusive(wq, &wait.wait, TASK_UNINTERRUPTIBLE);
501 if (xfs_isiflocked(ip))
502 io_schedule();
503 } while (!xfs_iflock_nowait(ip));
504
505 finish_wait(wq, &wait.wait);
506}
507
508STATIC uint
509_xfs_dic2xflags(
510 __uint16_t di_flags)
511{
512 uint flags = 0;
513
514 if (di_flags & XFS_DIFLAG_ANY) {
515 if (di_flags & XFS_DIFLAG_REALTIME)
516 flags |= XFS_XFLAG_REALTIME;
517 if (di_flags & XFS_DIFLAG_PREALLOC)
518 flags |= XFS_XFLAG_PREALLOC;
519 if (di_flags & XFS_DIFLAG_IMMUTABLE)
520 flags |= XFS_XFLAG_IMMUTABLE;
521 if (di_flags & XFS_DIFLAG_APPEND)
522 flags |= XFS_XFLAG_APPEND;
523 if (di_flags & XFS_DIFLAG_SYNC)
524 flags |= XFS_XFLAG_SYNC;
525 if (di_flags & XFS_DIFLAG_NOATIME)
526 flags |= XFS_XFLAG_NOATIME;
527 if (di_flags & XFS_DIFLAG_NODUMP)
528 flags |= XFS_XFLAG_NODUMP;
529 if (di_flags & XFS_DIFLAG_RTINHERIT)
530 flags |= XFS_XFLAG_RTINHERIT;
531 if (di_flags & XFS_DIFLAG_PROJINHERIT)
532 flags |= XFS_XFLAG_PROJINHERIT;
533 if (di_flags & XFS_DIFLAG_NOSYMLINKS)
534 flags |= XFS_XFLAG_NOSYMLINKS;
535 if (di_flags & XFS_DIFLAG_EXTSIZE)
536 flags |= XFS_XFLAG_EXTSIZE;
537 if (di_flags & XFS_DIFLAG_EXTSZINHERIT)
538 flags |= XFS_XFLAG_EXTSZINHERIT;
539 if (di_flags & XFS_DIFLAG_NODEFRAG)
540 flags |= XFS_XFLAG_NODEFRAG;
541 if (di_flags & XFS_DIFLAG_FILESTREAM)
542 flags |= XFS_XFLAG_FILESTREAM;
543 }
544
545 return flags;
546}
547
548uint
549xfs_ip2xflags(
550 xfs_inode_t *ip)
551{
552 xfs_icdinode_t *dic = &ip->i_d;
553
554 return _xfs_dic2xflags(dic->di_flags) |
555 (XFS_IFORK_Q(ip) ? XFS_XFLAG_HASATTR : 0);
556}
557
558uint
559xfs_dic2xflags(
560 xfs_dinode_t *dip)
561{
562 return _xfs_dic2xflags(be16_to_cpu(dip->di_flags)) |
563 (XFS_DFORK_Q(dip) ? XFS_XFLAG_HASATTR : 0);
564}
565
566/*
567 * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
568 * is allowed, otherwise it has to be an exact match. If a CI match is found,
569 * ci_name->name will point to a the actual name (caller must free) or
570 * will be set to NULL if an exact match is found.
571 */
572int
573xfs_lookup(
574 xfs_inode_t *dp,
575 struct xfs_name *name,
576 xfs_inode_t **ipp,
577 struct xfs_name *ci_name)
578{
579 xfs_ino_t inum;
580 int error;
581 uint lock_mode;
582
583 trace_xfs_lookup(dp, name);
584
585 if (XFS_FORCED_SHUTDOWN(dp->i_mount))
586 return XFS_ERROR(EIO);
587
588 lock_mode = xfs_ilock_data_map_shared(dp);
589 error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
590 xfs_iunlock(dp, lock_mode);
591
592 if (error)
593 goto out;
594
595 error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp);
596 if (error)
597 goto out_free_name;
598
599 return 0;
600
601out_free_name:
602 if (ci_name)
603 kmem_free(ci_name->name);
604out:
605 *ipp = NULL;
606 return error;
607}
608
609/*
610 * Allocate an inode on disk and return a copy of its in-core version.
611 * The in-core inode is locked exclusively. Set mode, nlink, and rdev
612 * appropriately within the inode. The uid and gid for the inode are
613 * set according to the contents of the given cred structure.
614 *
615 * Use xfs_dialloc() to allocate the on-disk inode. If xfs_dialloc()
616 * has a free inode available, call xfs_iget() to obtain the in-core
617 * version of the allocated inode. Finally, fill in the inode and
618 * log its initial contents. In this case, ialloc_context would be
619 * set to NULL.
620 *
621 * If xfs_dialloc() does not have an available inode, it will replenish
622 * its supply by doing an allocation. Since we can only do one
623 * allocation within a transaction without deadlocks, we must commit
624 * the current transaction before returning the inode itself.
625 * In this case, therefore, we will set ialloc_context and return.
626 * The caller should then commit the current transaction, start a new
627 * transaction, and call xfs_ialloc() again to actually get the inode.
628 *
629 * To ensure that some other process does not grab the inode that
630 * was allocated during the first call to xfs_ialloc(), this routine
631 * also returns the [locked] bp pointing to the head of the freelist
632 * as ialloc_context. The caller should hold this buffer across
633 * the commit and pass it back into this routine on the second call.
634 *
635 * If we are allocating quota inodes, we do not have a parent inode
636 * to attach to or associate with (i.e. pip == NULL) because they
637 * are not linked into the directory structure - they are attached
638 * directly to the superblock - and so have no parent.
639 */
640int
641xfs_ialloc(
642 xfs_trans_t *tp,
643 xfs_inode_t *pip,
644 umode_t mode,
645 xfs_nlink_t nlink,
646 xfs_dev_t rdev,
647 prid_t prid,
648 int okalloc,
649 xfs_buf_t **ialloc_context,
650 xfs_inode_t **ipp)
651{
652 struct xfs_mount *mp = tp->t_mountp;
653 xfs_ino_t ino;
654 xfs_inode_t *ip;
655 uint flags;
656 int error;
657 timespec_t tv;
658 int filestreams = 0;
659
660 /*
661 * Call the space management code to pick
662 * the on-disk inode to be allocated.
663 */
664 error = xfs_dialloc(tp, pip ? pip->i_ino : 0, mode, okalloc,
665 ialloc_context, &ino);
666 if (error)
667 return error;
668 if (*ialloc_context || ino == NULLFSINO) {
669 *ipp = NULL;
670 return 0;
671 }
672 ASSERT(*ialloc_context == NULL);
673
674 /*
675 * Get the in-core inode with the lock held exclusively.
676 * This is because we're setting fields here we need
677 * to prevent others from looking at until we're done.
678 */
679 error = xfs_iget(mp, tp, ino, XFS_IGET_CREATE,
680 XFS_ILOCK_EXCL, &ip);
681 if (error)
682 return error;
683 ASSERT(ip != NULL);
684
685 ip->i_d.di_mode = mode;
686 ip->i_d.di_onlink = 0;
687 ip->i_d.di_nlink = nlink;
688 ASSERT(ip->i_d.di_nlink == nlink);
689 ip->i_d.di_uid = xfs_kuid_to_uid(current_fsuid());
690 ip->i_d.di_gid = xfs_kgid_to_gid(current_fsgid());
691 xfs_set_projid(ip, prid);
692 memset(&(ip->i_d.di_pad[0]), 0, sizeof(ip->i_d.di_pad));
693
694 /*
695 * If the superblock version is up to where we support new format
696 * inodes and this is currently an old format inode, then change
697 * the inode version number now. This way we only do the conversion
698 * here rather than here and in the flush/logging code.
699 */
700 if (xfs_sb_version_hasnlink(&mp->m_sb) &&
701 ip->i_d.di_version == 1) {
702 ip->i_d.di_version = 2;
703 /*
704 * We've already zeroed the old link count, the projid field,
705 * and the pad field.
706 */
707 }
708
709 /*
710 * Project ids won't be stored on disk if we are using a version 1 inode.
711 */
712 if ((prid != 0) && (ip->i_d.di_version == 1))
713 xfs_bump_ino_vers2(tp, ip);
714
715 if (pip && XFS_INHERIT_GID(pip)) {
716 ip->i_d.di_gid = pip->i_d.di_gid;
717 if ((pip->i_d.di_mode & S_ISGID) && S_ISDIR(mode)) {
718 ip->i_d.di_mode |= S_ISGID;
719 }
720 }
721
722 /*
723 * If the group ID of the new file does not match the effective group
724 * ID or one of the supplementary group IDs, the S_ISGID bit is cleared
725 * (and only if the irix_sgid_inherit compatibility variable is set).
726 */
727 if ((irix_sgid_inherit) &&
728 (ip->i_d.di_mode & S_ISGID) &&
729 (!in_group_p(xfs_gid_to_kgid(ip->i_d.di_gid)))) {
730 ip->i_d.di_mode &= ~S_ISGID;
731 }
732
733 ip->i_d.di_size = 0;
734 ip->i_d.di_nextents = 0;
735 ASSERT(ip->i_d.di_nblocks == 0);
736
737 nanotime(&tv);
738 ip->i_d.di_mtime.t_sec = (__int32_t)tv.tv_sec;
739 ip->i_d.di_mtime.t_nsec = (__int32_t)tv.tv_nsec;
740 ip->i_d.di_atime = ip->i_d.di_mtime;
741 ip->i_d.di_ctime = ip->i_d.di_mtime;
742
743 /*
744 * di_gen will have been taken care of in xfs_iread.
745 */
746 ip->i_d.di_extsize = 0;
747 ip->i_d.di_dmevmask = 0;
748 ip->i_d.di_dmstate = 0;
749 ip->i_d.di_flags = 0;
750
751 if (ip->i_d.di_version == 3) {
752 ASSERT(ip->i_d.di_ino == ino);
753 ASSERT(uuid_equal(&ip->i_d.di_uuid, &mp->m_sb.sb_uuid));
754 ip->i_d.di_crc = 0;
755 ip->i_d.di_changecount = 1;
756 ip->i_d.di_lsn = 0;
757 ip->i_d.di_flags2 = 0;
758 memset(&(ip->i_d.di_pad2[0]), 0, sizeof(ip->i_d.di_pad2));
759 ip->i_d.di_crtime = ip->i_d.di_mtime;
760 }
761
762
763 flags = XFS_ILOG_CORE;
764 switch (mode & S_IFMT) {
765 case S_IFIFO:
766 case S_IFCHR:
767 case S_IFBLK:
768 case S_IFSOCK:
769 ip->i_d.di_format = XFS_DINODE_FMT_DEV;
770 ip->i_df.if_u2.if_rdev = rdev;
771 ip->i_df.if_flags = 0;
772 flags |= XFS_ILOG_DEV;
773 break;
774 case S_IFREG:
775 /*
776 * we can't set up filestreams until after the VFS inode
777 * is set up properly.
778 */
779 if (pip && xfs_inode_is_filestream(pip))
780 filestreams = 1;
781 /* fall through */
782 case S_IFDIR:
783 if (pip && (pip->i_d.di_flags & XFS_DIFLAG_ANY)) {
784 uint di_flags = 0;
785
786 if (S_ISDIR(mode)) {
787 if (pip->i_d.di_flags & XFS_DIFLAG_RTINHERIT)
788 di_flags |= XFS_DIFLAG_RTINHERIT;
789 if (pip->i_d.di_flags & XFS_DIFLAG_EXTSZINHERIT) {
790 di_flags |= XFS_DIFLAG_EXTSZINHERIT;
791 ip->i_d.di_extsize = pip->i_d.di_extsize;
792 }
793 } else if (S_ISREG(mode)) {
794 if (pip->i_d.di_flags & XFS_DIFLAG_RTINHERIT)
795 di_flags |= XFS_DIFLAG_REALTIME;
796 if (pip->i_d.di_flags & XFS_DIFLAG_EXTSZINHERIT) {
797 di_flags |= XFS_DIFLAG_EXTSIZE;
798 ip->i_d.di_extsize = pip->i_d.di_extsize;
799 }
800 }
801 if ((pip->i_d.di_flags & XFS_DIFLAG_NOATIME) &&
802 xfs_inherit_noatime)
803 di_flags |= XFS_DIFLAG_NOATIME;
804 if ((pip->i_d.di_flags & XFS_DIFLAG_NODUMP) &&
805 xfs_inherit_nodump)
806 di_flags |= XFS_DIFLAG_NODUMP;
807 if ((pip->i_d.di_flags & XFS_DIFLAG_SYNC) &&
808 xfs_inherit_sync)
809 di_flags |= XFS_DIFLAG_SYNC;
810 if ((pip->i_d.di_flags & XFS_DIFLAG_NOSYMLINKS) &&
811 xfs_inherit_nosymlinks)
812 di_flags |= XFS_DIFLAG_NOSYMLINKS;
813 if (pip->i_d.di_flags & XFS_DIFLAG_PROJINHERIT)
814 di_flags |= XFS_DIFLAG_PROJINHERIT;
815 if ((pip->i_d.di_flags & XFS_DIFLAG_NODEFRAG) &&
816 xfs_inherit_nodefrag)
817 di_flags |= XFS_DIFLAG_NODEFRAG;
818 if (pip->i_d.di_flags & XFS_DIFLAG_FILESTREAM)
819 di_flags |= XFS_DIFLAG_FILESTREAM;
820 ip->i_d.di_flags |= di_flags;
821 }
822 /* FALLTHROUGH */
823 case S_IFLNK:
824 ip->i_d.di_format = XFS_DINODE_FMT_EXTENTS;
825 ip->i_df.if_flags = XFS_IFEXTENTS;
826 ip->i_df.if_bytes = ip->i_df.if_real_bytes = 0;
827 ip->i_df.if_u1.if_extents = NULL;
828 break;
829 default:
830 ASSERT(0);
831 }
832 /*
833 * Attribute fork settings for new inode.
834 */
835 ip->i_d.di_aformat = XFS_DINODE_FMT_EXTENTS;
836 ip->i_d.di_anextents = 0;
837
838 /*
839 * Log the new values stuffed into the inode.
840 */
841 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
842 xfs_trans_log_inode(tp, ip, flags);
843
844 /* now that we have an i_mode we can setup inode ops and unlock */
845 xfs_setup_inode(ip);
846
847 /* now we have set up the vfs inode we can associate the filestream */
848 if (filestreams) {
849 error = xfs_filestream_associate(pip, ip);
850 if (error < 0)
851 return -error;
852 if (!error)
853 xfs_iflags_set(ip, XFS_IFILESTREAM);
854 }
855
856 *ipp = ip;
857 return 0;
858}
859
860/*
861 * Allocates a new inode from disk and return a pointer to the
862 * incore copy. This routine will internally commit the current
863 * transaction and allocate a new one if the Space Manager needed
864 * to do an allocation to replenish the inode free-list.
865 *
866 * This routine is designed to be called from xfs_create and
867 * xfs_create_dir.
868 *
869 */
870int
871xfs_dir_ialloc(
872 xfs_trans_t **tpp, /* input: current transaction;
873 output: may be a new transaction. */
874 xfs_inode_t *dp, /* directory within whose allocate
875 the inode. */
876 umode_t mode,
877 xfs_nlink_t nlink,
878 xfs_dev_t rdev,
879 prid_t prid, /* project id */
880 int okalloc, /* ok to allocate new space */
881 xfs_inode_t **ipp, /* pointer to inode; it will be
882 locked. */
883 int *committed)
884
885{
886 xfs_trans_t *tp;
887 xfs_trans_t *ntp;
888 xfs_inode_t *ip;
889 xfs_buf_t *ialloc_context = NULL;
890 int code;
891 void *dqinfo;
892 uint tflags;
893
894 tp = *tpp;
895 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
896
897 /*
898 * xfs_ialloc will return a pointer to an incore inode if
899 * the Space Manager has an available inode on the free
900 * list. Otherwise, it will do an allocation and replenish
901 * the freelist. Since we can only do one allocation per
902 * transaction without deadlocks, we will need to commit the
903 * current transaction and start a new one. We will then
904 * need to call xfs_ialloc again to get the inode.
905 *
906 * If xfs_ialloc did an allocation to replenish the freelist,
907 * it returns the bp containing the head of the freelist as
908 * ialloc_context. We will hold a lock on it across the
909 * transaction commit so that no other process can steal
910 * the inode(s) that we've just allocated.
911 */
912 code = xfs_ialloc(tp, dp, mode, nlink, rdev, prid, okalloc,
913 &ialloc_context, &ip);
914
915 /*
916 * Return an error if we were unable to allocate a new inode.
917 * This should only happen if we run out of space on disk or
918 * encounter a disk error.
919 */
920 if (code) {
921 *ipp = NULL;
922 return code;
923 }
924 if (!ialloc_context && !ip) {
925 *ipp = NULL;
926 return XFS_ERROR(ENOSPC);
927 }
928
929 /*
930 * If the AGI buffer is non-NULL, then we were unable to get an
931 * inode in one operation. We need to commit the current
932 * transaction and call xfs_ialloc() again. It is guaranteed
933 * to succeed the second time.
934 */
935 if (ialloc_context) {
936 struct xfs_trans_res tres;
937
938 /*
939 * Normally, xfs_trans_commit releases all the locks.
940 * We call bhold to hang on to the ialloc_context across
941 * the commit. Holding this buffer prevents any other
942 * processes from doing any allocations in this
943 * allocation group.
944 */
945 xfs_trans_bhold(tp, ialloc_context);
946 /*
947 * Save the log reservation so we can use
948 * them in the next transaction.
949 */
950 tres.tr_logres = xfs_trans_get_log_res(tp);
951 tres.tr_logcount = xfs_trans_get_log_count(tp);
952
953 /*
954 * We want the quota changes to be associated with the next
955 * transaction, NOT this one. So, detach the dqinfo from this
956 * and attach it to the next transaction.
957 */
958 dqinfo = NULL;
959 tflags = 0;
960 if (tp->t_dqinfo) {
961 dqinfo = (void *)tp->t_dqinfo;
962 tp->t_dqinfo = NULL;
963 tflags = tp->t_flags & XFS_TRANS_DQ_DIRTY;
964 tp->t_flags &= ~(XFS_TRANS_DQ_DIRTY);
965 }
966
967 ntp = xfs_trans_dup(tp);
968 code = xfs_trans_commit(tp, 0);
969 tp = ntp;
970 if (committed != NULL) {
971 *committed = 1;
972 }
973 /*
974 * If we get an error during the commit processing,
975 * release the buffer that is still held and return
976 * to the caller.
977 */
978 if (code) {
979 xfs_buf_relse(ialloc_context);
980 if (dqinfo) {
981 tp->t_dqinfo = dqinfo;
982 xfs_trans_free_dqinfo(tp);
983 }
984 *tpp = ntp;
985 *ipp = NULL;
986 return code;
987 }
988
989 /*
990 * transaction commit worked ok so we can drop the extra ticket
991 * reference that we gained in xfs_trans_dup()
992 */
993 xfs_log_ticket_put(tp->t_ticket);
994 tres.tr_logflags = XFS_TRANS_PERM_LOG_RES;
995 code = xfs_trans_reserve(tp, &tres, 0, 0);
996
997 /*
998 * Re-attach the quota info that we detached from prev trx.
999 */
1000 if (dqinfo) {
1001 tp->t_dqinfo = dqinfo;
1002 tp->t_flags |= tflags;
1003 }
1004
1005 if (code) {
1006 xfs_buf_relse(ialloc_context);
1007 *tpp = ntp;
1008 *ipp = NULL;
1009 return code;
1010 }
1011 xfs_trans_bjoin(tp, ialloc_context);
1012
1013 /*
1014 * Call ialloc again. Since we've locked out all
1015 * other allocations in this allocation group,
1016 * this call should always succeed.
1017 */
1018 code = xfs_ialloc(tp, dp, mode, nlink, rdev, prid,
1019 okalloc, &ialloc_context, &ip);
1020
1021 /*
1022 * If we get an error at this point, return to the caller
1023 * so that the current transaction can be aborted.
1024 */
1025 if (code) {
1026 *tpp = tp;
1027 *ipp = NULL;
1028 return code;
1029 }
1030 ASSERT(!ialloc_context && ip);
1031
1032 } else {
1033 if (committed != NULL)
1034 *committed = 0;
1035 }
1036
1037 *ipp = ip;
1038 *tpp = tp;
1039
1040 return 0;
1041}
1042
1043/*
1044 * Decrement the link count on an inode & log the change.
1045 * If this causes the link count to go to zero, initiate the
1046 * logging activity required to truncate a file.
1047 */
1048int /* error */
1049xfs_droplink(
1050 xfs_trans_t *tp,
1051 xfs_inode_t *ip)
1052{
1053 int error;
1054
1055 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
1056
1057 ASSERT (ip->i_d.di_nlink > 0);
1058 ip->i_d.di_nlink--;
1059 drop_nlink(VFS_I(ip));
1060 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1061
1062 error = 0;
1063 if (ip->i_d.di_nlink == 0) {
1064 /*
1065 * We're dropping the last link to this file.
1066 * Move the on-disk inode to the AGI unlinked list.
1067 * From xfs_inactive() we will pull the inode from
1068 * the list and free it.
1069 */
1070 error = xfs_iunlink(tp, ip);
1071 }
1072 return error;
1073}
1074
1075/*
1076 * This gets called when the inode's version needs to be changed from 1 to 2.
1077 * Currently this happens when the nlink field overflows the old 16-bit value
1078 * or when chproj is called to change the project for the first time.
1079 * As a side effect the superblock version will also get rev'd
1080 * to contain the NLINK bit.
1081 */
1082void
1083xfs_bump_ino_vers2(
1084 xfs_trans_t *tp,
1085 xfs_inode_t *ip)
1086{
1087 xfs_mount_t *mp;
1088
1089 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1090 ASSERT(ip->i_d.di_version == 1);
1091
1092 ip->i_d.di_version = 2;
1093 ip->i_d.di_onlink = 0;
1094 memset(&(ip->i_d.di_pad[0]), 0, sizeof(ip->i_d.di_pad));
1095 mp = tp->t_mountp;
1096 if (!xfs_sb_version_hasnlink(&mp->m_sb)) {
1097 spin_lock(&mp->m_sb_lock);
1098 if (!xfs_sb_version_hasnlink(&mp->m_sb)) {
1099 xfs_sb_version_addnlink(&mp->m_sb);
1100 spin_unlock(&mp->m_sb_lock);
1101 xfs_mod_sb(tp, XFS_SB_VERSIONNUM);
1102 } else {
1103 spin_unlock(&mp->m_sb_lock);
1104 }
1105 }
1106 /* Caller must log the inode */
1107}
1108
1109/*
1110 * Increment the link count on an inode & log the change.
1111 */
1112int
1113xfs_bumplink(
1114 xfs_trans_t *tp,
1115 xfs_inode_t *ip)
1116{
1117 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
1118
1119 ASSERT(ip->i_d.di_nlink > 0 || (VFS_I(ip)->i_state & I_LINKABLE));
1120 ip->i_d.di_nlink++;
1121 inc_nlink(VFS_I(ip));
1122 if ((ip->i_d.di_version == 1) &&
1123 (ip->i_d.di_nlink > XFS_MAXLINK_1)) {
1124 /*
1125 * The inode has increased its number of links beyond
1126 * what can fit in an old format inode. It now needs
1127 * to be converted to a version 2 inode with a 32 bit
1128 * link count. If this is the first inode in the file
1129 * system to do this, then we need to bump the superblock
1130 * version number as well.
1131 */
1132 xfs_bump_ino_vers2(tp, ip);
1133 }
1134
1135 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1136 return 0;
1137}
1138
1139int
1140xfs_create(
1141 xfs_inode_t *dp,
1142 struct xfs_name *name,
1143 umode_t mode,
1144 xfs_dev_t rdev,
1145 xfs_inode_t **ipp)
1146{
1147 int is_dir = S_ISDIR(mode);
1148 struct xfs_mount *mp = dp->i_mount;
1149 struct xfs_inode *ip = NULL;
1150 struct xfs_trans *tp = NULL;
1151 int error;
1152 xfs_bmap_free_t free_list;
1153 xfs_fsblock_t first_block;
1154 bool unlock_dp_on_error = false;
1155 uint cancel_flags;
1156 int committed;
1157 prid_t prid;
1158 struct xfs_dquot *udqp = NULL;
1159 struct xfs_dquot *gdqp = NULL;
1160 struct xfs_dquot *pdqp = NULL;
1161 struct xfs_trans_res tres;
1162 uint resblks;
1163
1164 trace_xfs_create(dp, name);
1165
1166 if (XFS_FORCED_SHUTDOWN(mp))
1167 return XFS_ERROR(EIO);
1168
1169 prid = xfs_get_initial_prid(dp);
1170
1171 /*
1172 * Make sure that we have allocated dquot(s) on disk.
1173 */
1174 error = xfs_qm_vop_dqalloc(dp, xfs_kuid_to_uid(current_fsuid()),
1175 xfs_kgid_to_gid(current_fsgid()), prid,
1176 XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
1177 &udqp, &gdqp, &pdqp);
1178 if (error)
1179 return error;
1180
1181 if (is_dir) {
1182 rdev = 0;
1183 resblks = XFS_MKDIR_SPACE_RES(mp, name->len);
1184 tres.tr_logres = M_RES(mp)->tr_mkdir.tr_logres;
1185 tres.tr_logcount = XFS_MKDIR_LOG_COUNT;
1186 tp = xfs_trans_alloc(mp, XFS_TRANS_MKDIR);
1187 } else {
1188 resblks = XFS_CREATE_SPACE_RES(mp, name->len);
1189 tres.tr_logres = M_RES(mp)->tr_create.tr_logres;
1190 tres.tr_logcount = XFS_CREATE_LOG_COUNT;
1191 tp = xfs_trans_alloc(mp, XFS_TRANS_CREATE);
1192 }
1193
1194 cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1195
1196 /*
1197 * Initially assume that the file does not exist and
1198 * reserve the resources for that case. If that is not
1199 * the case we'll drop the one we have and get a more
1200 * appropriate transaction later.
1201 */
1202 tres.tr_logflags = XFS_TRANS_PERM_LOG_RES;
1203 error = xfs_trans_reserve(tp, &tres, resblks, 0);
1204 if (error == ENOSPC) {
1205 /* flush outstanding delalloc blocks and retry */
1206 xfs_flush_inodes(mp);
1207 error = xfs_trans_reserve(tp, &tres, resblks, 0);
1208 }
1209 if (error == ENOSPC) {
1210 /* No space at all so try a "no-allocation" reservation */
1211 resblks = 0;
1212 error = xfs_trans_reserve(tp, &tres, 0, 0);
1213 }
1214 if (error) {
1215 cancel_flags = 0;
1216 goto out_trans_cancel;
1217 }
1218
1219 xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
1220 unlock_dp_on_error = true;
1221
1222 xfs_bmap_init(&free_list, &first_block);
1223
1224 /*
1225 * Reserve disk quota and the inode.
1226 */
1227 error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp,
1228 pdqp, resblks, 1, 0);
1229 if (error)
1230 goto out_trans_cancel;
1231
1232 error = xfs_dir_canenter(tp, dp, name, resblks);
1233 if (error)
1234 goto out_trans_cancel;
1235
1236 /*
1237 * A newly created regular or special file just has one directory
1238 * entry pointing to them, but a directory also the "." entry
1239 * pointing to itself.
1240 */
1241 error = xfs_dir_ialloc(&tp, dp, mode, is_dir ? 2 : 1, rdev,
1242 prid, resblks > 0, &ip, &committed);
1243 if (error) {
1244 if (error == ENOSPC)
1245 goto out_trans_cancel;
1246 goto out_trans_abort;
1247 }
1248
1249 /*
1250 * Now we join the directory inode to the transaction. We do not do it
1251 * earlier because xfs_dir_ialloc might commit the previous transaction
1252 * (and release all the locks). An error from here on will result in
1253 * the transaction cancel unlocking dp so don't do it explicitly in the
1254 * error path.
1255 */
1256 xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1257 unlock_dp_on_error = false;
1258
1259 error = xfs_dir_createname(tp, dp, name, ip->i_ino,
1260 &first_block, &free_list, resblks ?
1261 resblks - XFS_IALLOC_SPACE_RES(mp) : 0);
1262 if (error) {
1263 ASSERT(error != ENOSPC);
1264 goto out_trans_abort;
1265 }
1266 xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1267 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1268
1269 if (is_dir) {
1270 error = xfs_dir_init(tp, ip, dp);
1271 if (error)
1272 goto out_bmap_cancel;
1273
1274 error = xfs_bumplink(tp, dp);
1275 if (error)
1276 goto out_bmap_cancel;
1277 }
1278
1279 /*
1280 * If this is a synchronous mount, make sure that the
1281 * create transaction goes to disk before returning to
1282 * the user.
1283 */
1284 if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
1285 xfs_trans_set_sync(tp);
1286
1287 /*
1288 * Attach the dquot(s) to the inodes and modify them incore.
1289 * These ids of the inode couldn't have changed since the new
1290 * inode has been locked ever since it was created.
1291 */
1292 xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1293
1294 error = xfs_bmap_finish(&tp, &free_list, &committed);
1295 if (error)
1296 goto out_bmap_cancel;
1297
1298 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1299 if (error)
1300 goto out_release_inode;
1301
1302 xfs_qm_dqrele(udqp);
1303 xfs_qm_dqrele(gdqp);
1304 xfs_qm_dqrele(pdqp);
1305
1306 *ipp = ip;
1307 return 0;
1308
1309 out_bmap_cancel:
1310 xfs_bmap_cancel(&free_list);
1311 out_trans_abort:
1312 cancel_flags |= XFS_TRANS_ABORT;
1313 out_trans_cancel:
1314 xfs_trans_cancel(tp, cancel_flags);
1315 out_release_inode:
1316 /*
1317 * Wait until after the current transaction is aborted to
1318 * release the inode. This prevents recursive transactions
1319 * and deadlocks from xfs_inactive.
1320 */
1321 if (ip)
1322 IRELE(ip);
1323
1324 xfs_qm_dqrele(udqp);
1325 xfs_qm_dqrele(gdqp);
1326 xfs_qm_dqrele(pdqp);
1327
1328 if (unlock_dp_on_error)
1329 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1330 return error;
1331}
1332
1333int
1334xfs_create_tmpfile(
1335 struct xfs_inode *dp,
1336 struct dentry *dentry,
1337 umode_t mode,
1338 struct xfs_inode **ipp)
1339{
1340 struct xfs_mount *mp = dp->i_mount;
1341 struct xfs_inode *ip = NULL;
1342 struct xfs_trans *tp = NULL;
1343 int error;
1344 uint cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1345 prid_t prid;
1346 struct xfs_dquot *udqp = NULL;
1347 struct xfs_dquot *gdqp = NULL;
1348 struct xfs_dquot *pdqp = NULL;
1349 struct xfs_trans_res *tres;
1350 uint resblks;
1351
1352 if (XFS_FORCED_SHUTDOWN(mp))
1353 return XFS_ERROR(EIO);
1354
1355 prid = xfs_get_initial_prid(dp);
1356
1357 /*
1358 * Make sure that we have allocated dquot(s) on disk.
1359 */
1360 error = xfs_qm_vop_dqalloc(dp, xfs_kuid_to_uid(current_fsuid()),
1361 xfs_kgid_to_gid(current_fsgid()), prid,
1362 XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
1363 &udqp, &gdqp, &pdqp);
1364 if (error)
1365 return error;
1366
1367 resblks = XFS_IALLOC_SPACE_RES(mp);
1368 tp = xfs_trans_alloc(mp, XFS_TRANS_CREATE_TMPFILE);
1369
1370 tres = &M_RES(mp)->tr_create_tmpfile;
1371 error = xfs_trans_reserve(tp, tres, resblks, 0);
1372 if (error == ENOSPC) {
1373 /* No space at all so try a "no-allocation" reservation */
1374 resblks = 0;
1375 error = xfs_trans_reserve(tp, tres, 0, 0);
1376 }
1377 if (error) {
1378 cancel_flags = 0;
1379 goto out_trans_cancel;
1380 }
1381
1382 error = xfs_trans_reserve_quota(tp, mp, udqp, gdqp,
1383 pdqp, resblks, 1, 0);
1384 if (error)
1385 goto out_trans_cancel;
1386
1387 error = xfs_dir_ialloc(&tp, dp, mode, 1, 0,
1388 prid, resblks > 0, &ip, NULL);
1389 if (error) {
1390 if (error == ENOSPC)
1391 goto out_trans_cancel;
1392 goto out_trans_abort;
1393 }
1394
1395 if (mp->m_flags & XFS_MOUNT_WSYNC)
1396 xfs_trans_set_sync(tp);
1397
1398 /*
1399 * Attach the dquot(s) to the inodes and modify them incore.
1400 * These ids of the inode couldn't have changed since the new
1401 * inode has been locked ever since it was created.
1402 */
1403 xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1404
1405 ip->i_d.di_nlink--;
1406 error = xfs_iunlink(tp, ip);
1407 if (error)
1408 goto out_trans_abort;
1409
1410 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1411 if (error)
1412 goto out_release_inode;
1413
1414 xfs_qm_dqrele(udqp);
1415 xfs_qm_dqrele(gdqp);
1416 xfs_qm_dqrele(pdqp);
1417
1418 *ipp = ip;
1419 return 0;
1420
1421 out_trans_abort:
1422 cancel_flags |= XFS_TRANS_ABORT;
1423 out_trans_cancel:
1424 xfs_trans_cancel(tp, cancel_flags);
1425 out_release_inode:
1426 /*
1427 * Wait until after the current transaction is aborted to
1428 * release the inode. This prevents recursive transactions
1429 * and deadlocks from xfs_inactive.
1430 */
1431 if (ip)
1432 IRELE(ip);
1433
1434 xfs_qm_dqrele(udqp);
1435 xfs_qm_dqrele(gdqp);
1436 xfs_qm_dqrele(pdqp);
1437
1438 return error;
1439}
1440
1441int
1442xfs_link(
1443 xfs_inode_t *tdp,
1444 xfs_inode_t *sip,
1445 struct xfs_name *target_name)
1446{
1447 xfs_mount_t *mp = tdp->i_mount;
1448 xfs_trans_t *tp;
1449 int error;
1450 xfs_bmap_free_t free_list;
1451 xfs_fsblock_t first_block;
1452 int cancel_flags;
1453 int committed;
1454 int resblks;
1455
1456 trace_xfs_link(tdp, target_name);
1457
1458 ASSERT(!S_ISDIR(sip->i_d.di_mode));
1459
1460 if (XFS_FORCED_SHUTDOWN(mp))
1461 return XFS_ERROR(EIO);
1462
1463 error = xfs_qm_dqattach(sip, 0);
1464 if (error)
1465 goto std_return;
1466
1467 error = xfs_qm_dqattach(tdp, 0);
1468 if (error)
1469 goto std_return;
1470
1471 tp = xfs_trans_alloc(mp, XFS_TRANS_LINK);
1472 cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
1473 resblks = XFS_LINK_SPACE_RES(mp, target_name->len);
1474 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_link, resblks, 0);
1475 if (error == ENOSPC) {
1476 resblks = 0;
1477 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_link, 0, 0);
1478 }
1479 if (error) {
1480 cancel_flags = 0;
1481 goto error_return;
1482 }
1483
1484 xfs_lock_two_inodes(sip, tdp, XFS_ILOCK_EXCL);
1485
1486 xfs_trans_ijoin(tp, sip, XFS_ILOCK_EXCL);
1487 xfs_trans_ijoin(tp, tdp, XFS_ILOCK_EXCL);
1488
1489 /*
1490 * If we are using project inheritance, we only allow hard link
1491 * creation in our tree when the project IDs are the same; else
1492 * the tree quota mechanism could be circumvented.
1493 */
1494 if (unlikely((tdp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
1495 (xfs_get_projid(tdp) != xfs_get_projid(sip)))) {
1496 error = XFS_ERROR(EXDEV);
1497 goto error_return;
1498 }
1499
1500 error = xfs_dir_canenter(tp, tdp, target_name, resblks);
1501 if (error)
1502 goto error_return;
1503
1504 xfs_bmap_init(&free_list, &first_block);
1505
1506 if (sip->i_d.di_nlink == 0) {
1507 error = xfs_iunlink_remove(tp, sip);
1508 if (error)
1509 goto abort_return;
1510 }
1511
1512 error = xfs_dir_createname(tp, tdp, target_name, sip->i_ino,
1513 &first_block, &free_list, resblks);
1514 if (error)
1515 goto abort_return;
1516 xfs_trans_ichgtime(tp, tdp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1517 xfs_trans_log_inode(tp, tdp, XFS_ILOG_CORE);
1518
1519 error = xfs_bumplink(tp, sip);
1520 if (error)
1521 goto abort_return;
1522
1523 /*
1524 * If this is a synchronous mount, make sure that the
1525 * link transaction goes to disk before returning to
1526 * the user.
1527 */
1528 if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
1529 xfs_trans_set_sync(tp);
1530 }
1531
1532 error = xfs_bmap_finish (&tp, &free_list, &committed);
1533 if (error) {
1534 xfs_bmap_cancel(&free_list);
1535 goto abort_return;
1536 }
1537
1538 return xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1539
1540 abort_return:
1541 cancel_flags |= XFS_TRANS_ABORT;
1542 error_return:
1543 xfs_trans_cancel(tp, cancel_flags);
1544 std_return:
1545 return error;
1546}
1547
1548/*
1549 * Free up the underlying blocks past new_size. The new size must be smaller
1550 * than the current size. This routine can be used both for the attribute and
1551 * data fork, and does not modify the inode size, which is left to the caller.
1552 *
1553 * The transaction passed to this routine must have made a permanent log
1554 * reservation of at least XFS_ITRUNCATE_LOG_RES. This routine may commit the
1555 * given transaction and start new ones, so make sure everything involved in
1556 * the transaction is tidy before calling here. Some transaction will be
1557 * returned to the caller to be committed. The incoming transaction must
1558 * already include the inode, and both inode locks must be held exclusively.
1559 * The inode must also be "held" within the transaction. On return the inode
1560 * will be "held" within the returned transaction. This routine does NOT
1561 * require any disk space to be reserved for it within the transaction.
1562 *
1563 * If we get an error, we must return with the inode locked and linked into the
1564 * current transaction. This keeps things simple for the higher level code,
1565 * because it always knows that the inode is locked and held in the transaction
1566 * that returns to it whether errors occur or not. We don't mark the inode
1567 * dirty on error so that transactions can be easily aborted if possible.
1568 */
1569int
1570xfs_itruncate_extents(
1571 struct xfs_trans **tpp,
1572 struct xfs_inode *ip,
1573 int whichfork,
1574 xfs_fsize_t new_size)
1575{
1576 struct xfs_mount *mp = ip->i_mount;
1577 struct xfs_trans *tp = *tpp;
1578 struct xfs_trans *ntp;
1579 xfs_bmap_free_t free_list;
1580 xfs_fsblock_t first_block;
1581 xfs_fileoff_t first_unmap_block;
1582 xfs_fileoff_t last_block;
1583 xfs_filblks_t unmap_len;
1584 int committed;
1585 int error = 0;
1586 int done = 0;
1587
1588 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1589 ASSERT(!atomic_read(&VFS_I(ip)->i_count) ||
1590 xfs_isilocked(ip, XFS_IOLOCK_EXCL));
1591 ASSERT(new_size <= XFS_ISIZE(ip));
1592 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
1593 ASSERT(ip->i_itemp != NULL);
1594 ASSERT(ip->i_itemp->ili_lock_flags == 0);
1595 ASSERT(!XFS_NOT_DQATTACHED(mp, ip));
1596
1597 trace_xfs_itruncate_extents_start(ip, new_size);
1598
1599 /*
1600 * Since it is possible for space to become allocated beyond
1601 * the end of the file (in a crash where the space is allocated
1602 * but the inode size is not yet updated), simply remove any
1603 * blocks which show up between the new EOF and the maximum
1604 * possible file size. If the first block to be removed is
1605 * beyond the maximum file size (ie it is the same as last_block),
1606 * then there is nothing to do.
1607 */
1608 first_unmap_block = XFS_B_TO_FSB(mp, (xfs_ufsize_t)new_size);
1609 last_block = XFS_B_TO_FSB(mp, mp->m_super->s_maxbytes);
1610 if (first_unmap_block == last_block)
1611 return 0;
1612
1613 ASSERT(first_unmap_block < last_block);
1614 unmap_len = last_block - first_unmap_block + 1;
1615 while (!done) {
1616 xfs_bmap_init(&free_list, &first_block);
1617 error = xfs_bunmapi(tp, ip,
1618 first_unmap_block, unmap_len,
1619 xfs_bmapi_aflag(whichfork),
1620 XFS_ITRUNC_MAX_EXTENTS,
1621 &first_block, &free_list,
1622 &done);
1623 if (error)
1624 goto out_bmap_cancel;
1625
1626 /*
1627 * Duplicate the transaction that has the permanent
1628 * reservation and commit the old transaction.
1629 */
1630 error = xfs_bmap_finish(&tp, &free_list, &committed);
1631 if (committed)
1632 xfs_trans_ijoin(tp, ip, 0);
1633 if (error)
1634 goto out_bmap_cancel;
1635
1636 if (committed) {
1637 /*
1638 * Mark the inode dirty so it will be logged and
1639 * moved forward in the log as part of every commit.
1640 */
1641 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1642 }
1643
1644 ntp = xfs_trans_dup(tp);
1645 error = xfs_trans_commit(tp, 0);
1646 tp = ntp;
1647
1648 xfs_trans_ijoin(tp, ip, 0);
1649
1650 if (error)
1651 goto out;
1652
1653 /*
1654 * Transaction commit worked ok so we can drop the extra ticket
1655 * reference that we gained in xfs_trans_dup()
1656 */
1657 xfs_log_ticket_put(tp->t_ticket);
1658 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_itruncate, 0, 0);
1659 if (error)
1660 goto out;
1661 }
1662
1663 /*
1664 * Always re-log the inode so that our permanent transaction can keep
1665 * on rolling it forward in the log.
1666 */
1667 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1668
1669 trace_xfs_itruncate_extents_end(ip, new_size);
1670
1671out:
1672 *tpp = tp;
1673 return error;
1674out_bmap_cancel:
1675 /*
1676 * If the bunmapi call encounters an error, return to the caller where
1677 * the transaction can be properly aborted. We just need to make sure
1678 * we're not holding any resources that we were not when we came in.
1679 */
1680 xfs_bmap_cancel(&free_list);
1681 goto out;
1682}
1683
1684int
1685xfs_release(
1686 xfs_inode_t *ip)
1687{
1688 xfs_mount_t *mp = ip->i_mount;
1689 int error;
1690
1691 if (!S_ISREG(ip->i_d.di_mode) || (ip->i_d.di_mode == 0))
1692 return 0;
1693
1694 /* If this is a read-only mount, don't do this (would generate I/O) */
1695 if (mp->m_flags & XFS_MOUNT_RDONLY)
1696 return 0;
1697
1698 if (!XFS_FORCED_SHUTDOWN(mp)) {
1699 int truncated;
1700
1701 /*
1702 * If we are using filestreams, and we have an unlinked
1703 * file that we are processing the last close on, then nothing
1704 * will be able to reopen and write to this file. Purge this
1705 * inode from the filestreams cache so that it doesn't delay
1706 * teardown of the inode.
1707 */
1708 if ((ip->i_d.di_nlink == 0) && xfs_inode_is_filestream(ip))
1709 xfs_filestream_deassociate(ip);
1710
1711 /*
1712 * If we previously truncated this file and removed old data
1713 * in the process, we want to initiate "early" writeout on
1714 * the last close. This is an attempt to combat the notorious
1715 * NULL files problem which is particularly noticeable from a
1716 * truncate down, buffered (re-)write (delalloc), followed by
1717 * a crash. What we are effectively doing here is
1718 * significantly reducing the time window where we'd otherwise
1719 * be exposed to that problem.
1720 */
1721 truncated = xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED);
1722 if (truncated) {
1723 xfs_iflags_clear(ip, XFS_IDIRTY_RELEASE);
1724 if (VN_DIRTY(VFS_I(ip)) && ip->i_delayed_blks > 0) {
1725 error = -filemap_flush(VFS_I(ip)->i_mapping);
1726 if (error)
1727 return error;
1728 }
1729 }
1730 }
1731
1732 if (ip->i_d.di_nlink == 0)
1733 return 0;
1734
1735 if (xfs_can_free_eofblocks(ip, false)) {
1736
1737 /*
1738 * If we can't get the iolock just skip truncating the blocks
1739 * past EOF because we could deadlock with the mmap_sem
1740 * otherwise. We'll get another chance to drop them once the
1741 * last reference to the inode is dropped, so we'll never leak
1742 * blocks permanently.
1743 *
1744 * Further, check if the inode is being opened, written and
1745 * closed frequently and we have delayed allocation blocks
1746 * outstanding (e.g. streaming writes from the NFS server),
1747 * truncating the blocks past EOF will cause fragmentation to
1748 * occur.
1749 *
1750 * In this case don't do the truncation, either, but we have to
1751 * be careful how we detect this case. Blocks beyond EOF show
1752 * up as i_delayed_blks even when the inode is clean, so we
1753 * need to truncate them away first before checking for a dirty
1754 * release. Hence on the first dirty close we will still remove
1755 * the speculative allocation, but after that we will leave it
1756 * in place.
1757 */
1758 if (xfs_iflags_test(ip, XFS_IDIRTY_RELEASE))
1759 return 0;
1760
1761 error = xfs_free_eofblocks(mp, ip, true);
1762 if (error && error != EAGAIN)
1763 return error;
1764
1765 /* delalloc blocks after truncation means it really is dirty */
1766 if (ip->i_delayed_blks)
1767 xfs_iflags_set(ip, XFS_IDIRTY_RELEASE);
1768 }
1769 return 0;
1770}
1771
1772/*
1773 * xfs_inactive_truncate
1774 *
1775 * Called to perform a truncate when an inode becomes unlinked.
1776 */
1777STATIC int
1778xfs_inactive_truncate(
1779 struct xfs_inode *ip)
1780{
1781 struct xfs_mount *mp = ip->i_mount;
1782 struct xfs_trans *tp;
1783 int error;
1784
1785 tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
1786 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_itruncate, 0, 0);
1787 if (error) {
1788 ASSERT(XFS_FORCED_SHUTDOWN(mp));
1789 xfs_trans_cancel(tp, 0);
1790 return error;
1791 }
1792
1793 xfs_ilock(ip, XFS_ILOCK_EXCL);
1794 xfs_trans_ijoin(tp, ip, 0);
1795
1796 /*
1797 * Log the inode size first to prevent stale data exposure in the event
1798 * of a system crash before the truncate completes. See the related
1799 * comment in xfs_setattr_size() for details.
1800 */
1801 ip->i_d.di_size = 0;
1802 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1803
1804 error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK, 0);
1805 if (error)
1806 goto error_trans_cancel;
1807
1808 ASSERT(ip->i_d.di_nextents == 0);
1809
1810 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1811 if (error)
1812 goto error_unlock;
1813
1814 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1815 return 0;
1816
1817error_trans_cancel:
1818 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES | XFS_TRANS_ABORT);
1819error_unlock:
1820 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1821 return error;
1822}
1823
1824/*
1825 * xfs_inactive_ifree()
1826 *
1827 * Perform the inode free when an inode is unlinked.
1828 */
1829STATIC int
1830xfs_inactive_ifree(
1831 struct xfs_inode *ip)
1832{
1833 xfs_bmap_free_t free_list;
1834 xfs_fsblock_t first_block;
1835 int committed;
1836 struct xfs_mount *mp = ip->i_mount;
1837 struct xfs_trans *tp;
1838 int error;
1839
1840 tp = xfs_trans_alloc(mp, XFS_TRANS_INACTIVE);
1841 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_ifree, 0, 0);
1842 if (error) {
1843 ASSERT(XFS_FORCED_SHUTDOWN(mp));
1844 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES);
1845 return error;
1846 }
1847
1848 xfs_ilock(ip, XFS_ILOCK_EXCL);
1849 xfs_trans_ijoin(tp, ip, 0);
1850
1851 xfs_bmap_init(&free_list, &first_block);
1852 error = xfs_ifree(tp, ip, &free_list);
1853 if (error) {
1854 /*
1855 * If we fail to free the inode, shut down. The cancel
1856 * might do that, we need to make sure. Otherwise the
1857 * inode might be lost for a long time or forever.
1858 */
1859 if (!XFS_FORCED_SHUTDOWN(mp)) {
1860 xfs_notice(mp, "%s: xfs_ifree returned error %d",
1861 __func__, error);
1862 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1863 }
1864 xfs_trans_cancel(tp, XFS_TRANS_RELEASE_LOG_RES|XFS_TRANS_ABORT);
1865 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1866 return error;
1867 }
1868
1869 /*
1870 * Credit the quota account(s). The inode is gone.
1871 */
1872 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
1873
1874 /*
1875 * Just ignore errors at this point. There is nothing we can
1876 * do except to try to keep going. Make sure it's not a silent
1877 * error.
1878 */
1879 error = xfs_bmap_finish(&tp, &free_list, &committed);
1880 if (error)
1881 xfs_notice(mp, "%s: xfs_bmap_finish returned error %d",
1882 __func__, error);
1883 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
1884 if (error)
1885 xfs_notice(mp, "%s: xfs_trans_commit returned error %d",
1886 __func__, error);
1887
1888 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1889 return 0;
1890}
1891
1892/*
1893 * xfs_inactive
1894 *
1895 * This is called when the vnode reference count for the vnode
1896 * goes to zero. If the file has been unlinked, then it must
1897 * now be truncated. Also, we clear all of the read-ahead state
1898 * kept for the inode here since the file is now closed.
1899 */
1900void
1901xfs_inactive(
1902 xfs_inode_t *ip)
1903{
1904 struct xfs_mount *mp;
1905 int error;
1906 int truncate = 0;
1907
1908 /*
1909 * If the inode is already free, then there can be nothing
1910 * to clean up here.
1911 */
1912 if (ip->i_d.di_mode == 0) {
1913 ASSERT(ip->i_df.if_real_bytes == 0);
1914 ASSERT(ip->i_df.if_broot_bytes == 0);
1915 return;
1916 }
1917
1918 mp = ip->i_mount;
1919
1920 /* If this is a read-only mount, don't do this (would generate I/O) */
1921 if (mp->m_flags & XFS_MOUNT_RDONLY)
1922 return;
1923
1924 if (ip->i_d.di_nlink != 0) {
1925 /*
1926 * force is true because we are evicting an inode from the
1927 * cache. Post-eof blocks must be freed, lest we end up with
1928 * broken free space accounting.
1929 */
1930 if (xfs_can_free_eofblocks(ip, true))
1931 xfs_free_eofblocks(mp, ip, false);
1932
1933 return;
1934 }
1935
1936 if (S_ISREG(ip->i_d.di_mode) &&
1937 (ip->i_d.di_size != 0 || XFS_ISIZE(ip) != 0 ||
1938 ip->i_d.di_nextents > 0 || ip->i_delayed_blks > 0))
1939 truncate = 1;
1940
1941 error = xfs_qm_dqattach(ip, 0);
1942 if (error)
1943 return;
1944
1945 if (S_ISLNK(ip->i_d.di_mode))
1946 error = xfs_inactive_symlink(ip);
1947 else if (truncate)
1948 error = xfs_inactive_truncate(ip);
1949 if (error)
1950 return;
1951
1952 /*
1953 * If there are attributes associated with the file then blow them away
1954 * now. The code calls a routine that recursively deconstructs the
1955 * attribute fork. We need to just commit the current transaction
1956 * because we can't use it for xfs_attr_inactive().
1957 */
1958 if (ip->i_d.di_anextents > 0) {
1959 ASSERT(ip->i_d.di_forkoff != 0);
1960
1961 error = xfs_attr_inactive(ip);
1962 if (error)
1963 return;
1964 }
1965
1966 if (ip->i_afp)
1967 xfs_idestroy_fork(ip, XFS_ATTR_FORK);
1968
1969 ASSERT(ip->i_d.di_anextents == 0);
1970
1971 /*
1972 * Free the inode.
1973 */
1974 error = xfs_inactive_ifree(ip);
1975 if (error)
1976 return;
1977
1978 /*
1979 * Release the dquots held by inode, if any.
1980 */
1981 xfs_qm_dqdetach(ip);
1982}
1983
1984/*
1985 * This is called when the inode's link count goes to 0.
1986 * We place the on-disk inode on a list in the AGI. It
1987 * will be pulled from this list when the inode is freed.
1988 */
1989int
1990xfs_iunlink(
1991 xfs_trans_t *tp,
1992 xfs_inode_t *ip)
1993{
1994 xfs_mount_t *mp;
1995 xfs_agi_t *agi;
1996 xfs_dinode_t *dip;
1997 xfs_buf_t *agibp;
1998 xfs_buf_t *ibp;
1999 xfs_agino_t agino;
2000 short bucket_index;
2001 int offset;
2002 int error;
2003
2004 ASSERT(ip->i_d.di_nlink == 0);
2005 ASSERT(ip->i_d.di_mode != 0);
2006
2007 mp = tp->t_mountp;
2008
2009 /*
2010 * Get the agi buffer first. It ensures lock ordering
2011 * on the list.
2012 */
2013 error = xfs_read_agi(mp, tp, XFS_INO_TO_AGNO(mp, ip->i_ino), &agibp);
2014 if (error)
2015 return error;
2016 agi = XFS_BUF_TO_AGI(agibp);
2017
2018 /*
2019 * Get the index into the agi hash table for the
2020 * list this inode will go on.
2021 */
2022 agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2023 ASSERT(agino != 0);
2024 bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
2025 ASSERT(agi->agi_unlinked[bucket_index]);
2026 ASSERT(be32_to_cpu(agi->agi_unlinked[bucket_index]) != agino);
2027
2028 if (agi->agi_unlinked[bucket_index] != cpu_to_be32(NULLAGINO)) {
2029 /*
2030 * There is already another inode in the bucket we need
2031 * to add ourselves to. Add us at the front of the list.
2032 * Here we put the head pointer into our next pointer,
2033 * and then we fall through to point the head at us.
2034 */
2035 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &dip, &ibp,
2036 0, 0);
2037 if (error)
2038 return error;
2039
2040 ASSERT(dip->di_next_unlinked == cpu_to_be32(NULLAGINO));
2041 dip->di_next_unlinked = agi->agi_unlinked[bucket_index];
2042 offset = ip->i_imap.im_boffset +
2043 offsetof(xfs_dinode_t, di_next_unlinked);
2044
2045 /* need to recalc the inode CRC if appropriate */
2046 xfs_dinode_calc_crc(mp, dip);
2047
2048 xfs_trans_inode_buf(tp, ibp);
2049 xfs_trans_log_buf(tp, ibp, offset,
2050 (offset + sizeof(xfs_agino_t) - 1));
2051 xfs_inobp_check(mp, ibp);
2052 }
2053
2054 /*
2055 * Point the bucket head pointer at the inode being inserted.
2056 */
2057 ASSERT(agino != 0);
2058 agi->agi_unlinked[bucket_index] = cpu_to_be32(agino);
2059 offset = offsetof(xfs_agi_t, agi_unlinked) +
2060 (sizeof(xfs_agino_t) * bucket_index);
2061 xfs_trans_log_buf(tp, agibp, offset,
2062 (offset + sizeof(xfs_agino_t) - 1));
2063 return 0;
2064}
2065
2066/*
2067 * Pull the on-disk inode from the AGI unlinked list.
2068 */
2069STATIC int
2070xfs_iunlink_remove(
2071 xfs_trans_t *tp,
2072 xfs_inode_t *ip)
2073{
2074 xfs_ino_t next_ino;
2075 xfs_mount_t *mp;
2076 xfs_agi_t *agi;
2077 xfs_dinode_t *dip;
2078 xfs_buf_t *agibp;
2079 xfs_buf_t *ibp;
2080 xfs_agnumber_t agno;
2081 xfs_agino_t agino;
2082 xfs_agino_t next_agino;
2083 xfs_buf_t *last_ibp;
2084 xfs_dinode_t *last_dip = NULL;
2085 short bucket_index;
2086 int offset, last_offset = 0;
2087 int error;
2088
2089 mp = tp->t_mountp;
2090 agno = XFS_INO_TO_AGNO(mp, ip->i_ino);
2091
2092 /*
2093 * Get the agi buffer first. It ensures lock ordering
2094 * on the list.
2095 */
2096 error = xfs_read_agi(mp, tp, agno, &agibp);
2097 if (error)
2098 return error;
2099
2100 agi = XFS_BUF_TO_AGI(agibp);
2101
2102 /*
2103 * Get the index into the agi hash table for the
2104 * list this inode will go on.
2105 */
2106 agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2107 ASSERT(agino != 0);
2108 bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
2109 ASSERT(agi->agi_unlinked[bucket_index] != cpu_to_be32(NULLAGINO));
2110 ASSERT(agi->agi_unlinked[bucket_index]);
2111
2112 if (be32_to_cpu(agi->agi_unlinked[bucket_index]) == agino) {
2113 /*
2114 * We're at the head of the list. Get the inode's on-disk
2115 * buffer to see if there is anyone after us on the list.
2116 * Only modify our next pointer if it is not already NULLAGINO.
2117 * This saves us the overhead of dealing with the buffer when
2118 * there is no need to change it.
2119 */
2120 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &dip, &ibp,
2121 0, 0);
2122 if (error) {
2123 xfs_warn(mp, "%s: xfs_imap_to_bp returned error %d.",
2124 __func__, error);
2125 return error;
2126 }
2127 next_agino = be32_to_cpu(dip->di_next_unlinked);
2128 ASSERT(next_agino != 0);
2129 if (next_agino != NULLAGINO) {
2130 dip->di_next_unlinked = cpu_to_be32(NULLAGINO);
2131 offset = ip->i_imap.im_boffset +
2132 offsetof(xfs_dinode_t, di_next_unlinked);
2133
2134 /* need to recalc the inode CRC if appropriate */
2135 xfs_dinode_calc_crc(mp, dip);
2136
2137 xfs_trans_inode_buf(tp, ibp);
2138 xfs_trans_log_buf(tp, ibp, offset,
2139 (offset + sizeof(xfs_agino_t) - 1));
2140 xfs_inobp_check(mp, ibp);
2141 } else {
2142 xfs_trans_brelse(tp, ibp);
2143 }
2144 /*
2145 * Point the bucket head pointer at the next inode.
2146 */
2147 ASSERT(next_agino != 0);
2148 ASSERT(next_agino != agino);
2149 agi->agi_unlinked[bucket_index] = cpu_to_be32(next_agino);
2150 offset = offsetof(xfs_agi_t, agi_unlinked) +
2151 (sizeof(xfs_agino_t) * bucket_index);
2152 xfs_trans_log_buf(tp, agibp, offset,
2153 (offset + sizeof(xfs_agino_t) - 1));
2154 } else {
2155 /*
2156 * We need to search the list for the inode being freed.
2157 */
2158 next_agino = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2159 last_ibp = NULL;
2160 while (next_agino != agino) {
2161 struct xfs_imap imap;
2162
2163 if (last_ibp)
2164 xfs_trans_brelse(tp, last_ibp);
2165
2166 imap.im_blkno = 0;
2167 next_ino = XFS_AGINO_TO_INO(mp, agno, next_agino);
2168
2169 error = xfs_imap(mp, tp, next_ino, &imap, 0);
2170 if (error) {
2171 xfs_warn(mp,
2172 "%s: xfs_imap returned error %d.",
2173 __func__, error);
2174 return error;
2175 }
2176
2177 error = xfs_imap_to_bp(mp, tp, &imap, &last_dip,
2178 &last_ibp, 0, 0);
2179 if (error) {
2180 xfs_warn(mp,
2181 "%s: xfs_imap_to_bp returned error %d.",
2182 __func__, error);
2183 return error;
2184 }
2185
2186 last_offset = imap.im_boffset;
2187 next_agino = be32_to_cpu(last_dip->di_next_unlinked);
2188 ASSERT(next_agino != NULLAGINO);
2189 ASSERT(next_agino != 0);
2190 }
2191
2192 /*
2193 * Now last_ibp points to the buffer previous to us on the
2194 * unlinked list. Pull us from the list.
2195 */
2196 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &dip, &ibp,
2197 0, 0);
2198 if (error) {
2199 xfs_warn(mp, "%s: xfs_imap_to_bp(2) returned error %d.",
2200 __func__, error);
2201 return error;
2202 }
2203 next_agino = be32_to_cpu(dip->di_next_unlinked);
2204 ASSERT(next_agino != 0);
2205 ASSERT(next_agino != agino);
2206 if (next_agino != NULLAGINO) {
2207 dip->di_next_unlinked = cpu_to_be32(NULLAGINO);
2208 offset = ip->i_imap.im_boffset +
2209 offsetof(xfs_dinode_t, di_next_unlinked);
2210
2211 /* need to recalc the inode CRC if appropriate */
2212 xfs_dinode_calc_crc(mp, dip);
2213
2214 xfs_trans_inode_buf(tp, ibp);
2215 xfs_trans_log_buf(tp, ibp, offset,
2216 (offset + sizeof(xfs_agino_t) - 1));
2217 xfs_inobp_check(mp, ibp);
2218 } else {
2219 xfs_trans_brelse(tp, ibp);
2220 }
2221 /*
2222 * Point the previous inode on the list to the next inode.
2223 */
2224 last_dip->di_next_unlinked = cpu_to_be32(next_agino);
2225 ASSERT(next_agino != 0);
2226 offset = last_offset + offsetof(xfs_dinode_t, di_next_unlinked);
2227
2228 /* need to recalc the inode CRC if appropriate */
2229 xfs_dinode_calc_crc(mp, last_dip);
2230
2231 xfs_trans_inode_buf(tp, last_ibp);
2232 xfs_trans_log_buf(tp, last_ibp, offset,
2233 (offset + sizeof(xfs_agino_t) - 1));
2234 xfs_inobp_check(mp, last_ibp);
2235 }
2236 return 0;
2237}
2238
2239/*
2240 * A big issue when freeing the inode cluster is that we _cannot_ skip any
2241 * inodes that are in memory - they all must be marked stale and attached to
2242 * the cluster buffer.
2243 */
2244STATIC int
2245xfs_ifree_cluster(
2246 xfs_inode_t *free_ip,
2247 xfs_trans_t *tp,
2248 xfs_ino_t inum)
2249{
2250 xfs_mount_t *mp = free_ip->i_mount;
2251 int blks_per_cluster;
2252 int inodes_per_cluster;
2253 int nbufs;
2254 int i, j;
2255 xfs_daddr_t blkno;
2256 xfs_buf_t *bp;
2257 xfs_inode_t *ip;
2258 xfs_inode_log_item_t *iip;
2259 xfs_log_item_t *lip;
2260 struct xfs_perag *pag;
2261
2262 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, inum));
2263 blks_per_cluster = xfs_icluster_size_fsb(mp);
2264 inodes_per_cluster = blks_per_cluster << mp->m_sb.sb_inopblog;
2265 nbufs = mp->m_ialloc_blks / blks_per_cluster;
2266
2267 for (j = 0; j < nbufs; j++, inum += inodes_per_cluster) {
2268 blkno = XFS_AGB_TO_DADDR(mp, XFS_INO_TO_AGNO(mp, inum),
2269 XFS_INO_TO_AGBNO(mp, inum));
2270
2271 /*
2272 * We obtain and lock the backing buffer first in the process
2273 * here, as we have to ensure that any dirty inode that we
2274 * can't get the flush lock on is attached to the buffer.
2275 * If we scan the in-memory inodes first, then buffer IO can
2276 * complete before we get a lock on it, and hence we may fail
2277 * to mark all the active inodes on the buffer stale.
2278 */
2279 bp = xfs_trans_get_buf(tp, mp->m_ddev_targp, blkno,
2280 mp->m_bsize * blks_per_cluster,
2281 XBF_UNMAPPED);
2282
2283 if (!bp)
2284 return ENOMEM;
2285
2286 /*
2287 * This buffer may not have been correctly initialised as we
2288 * didn't read it from disk. That's not important because we are
2289 * only using to mark the buffer as stale in the log, and to
2290 * attach stale cached inodes on it. That means it will never be
2291 * dispatched for IO. If it is, we want to know about it, and we
2292 * want it to fail. We can acheive this by adding a write
2293 * verifier to the buffer.
2294 */
2295 bp->b_ops = &xfs_inode_buf_ops;
2296
2297 /*
2298 * Walk the inodes already attached to the buffer and mark them
2299 * stale. These will all have the flush locks held, so an
2300 * in-memory inode walk can't lock them. By marking them all
2301 * stale first, we will not attempt to lock them in the loop
2302 * below as the XFS_ISTALE flag will be set.
2303 */
2304 lip = bp->b_fspriv;
2305 while (lip) {
2306 if (lip->li_type == XFS_LI_INODE) {
2307 iip = (xfs_inode_log_item_t *)lip;
2308 ASSERT(iip->ili_logged == 1);
2309 lip->li_cb = xfs_istale_done;
2310 xfs_trans_ail_copy_lsn(mp->m_ail,
2311 &iip->ili_flush_lsn,
2312 &iip->ili_item.li_lsn);
2313 xfs_iflags_set(iip->ili_inode, XFS_ISTALE);
2314 }
2315 lip = lip->li_bio_list;
2316 }
2317
2318
2319 /*
2320 * For each inode in memory attempt to add it to the inode
2321 * buffer and set it up for being staled on buffer IO
2322 * completion. This is safe as we've locked out tail pushing
2323 * and flushing by locking the buffer.
2324 *
2325 * We have already marked every inode that was part of a
2326 * transaction stale above, which means there is no point in
2327 * even trying to lock them.
2328 */
2329 for (i = 0; i < inodes_per_cluster; i++) {
2330retry:
2331 rcu_read_lock();
2332 ip = radix_tree_lookup(&pag->pag_ici_root,
2333 XFS_INO_TO_AGINO(mp, (inum + i)));
2334
2335 /* Inode not in memory, nothing to do */
2336 if (!ip) {
2337 rcu_read_unlock();
2338 continue;
2339 }
2340
2341 /*
2342 * because this is an RCU protected lookup, we could
2343 * find a recently freed or even reallocated inode
2344 * during the lookup. We need to check under the
2345 * i_flags_lock for a valid inode here. Skip it if it
2346 * is not valid, the wrong inode or stale.
2347 */
2348 spin_lock(&ip->i_flags_lock);
2349 if (ip->i_ino != inum + i ||
2350 __xfs_iflags_test(ip, XFS_ISTALE)) {
2351 spin_unlock(&ip->i_flags_lock);
2352 rcu_read_unlock();
2353 continue;
2354 }
2355 spin_unlock(&ip->i_flags_lock);
2356
2357 /*
2358 * Don't try to lock/unlock the current inode, but we
2359 * _cannot_ skip the other inodes that we did not find
2360 * in the list attached to the buffer and are not
2361 * already marked stale. If we can't lock it, back off
2362 * and retry.
2363 */
2364 if (ip != free_ip &&
2365 !xfs_ilock_nowait(ip, XFS_ILOCK_EXCL)) {
2366 rcu_read_unlock();
2367 delay(1);
2368 goto retry;
2369 }
2370 rcu_read_unlock();
2371
2372 xfs_iflock(ip);
2373 xfs_iflags_set(ip, XFS_ISTALE);
2374
2375 /*
2376 * we don't need to attach clean inodes or those only
2377 * with unlogged changes (which we throw away, anyway).
2378 */
2379 iip = ip->i_itemp;
2380 if (!iip || xfs_inode_clean(ip)) {
2381 ASSERT(ip != free_ip);
2382 xfs_ifunlock(ip);
2383 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2384 continue;
2385 }
2386
2387 iip->ili_last_fields = iip->ili_fields;
2388 iip->ili_fields = 0;
2389 iip->ili_logged = 1;
2390 xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
2391 &iip->ili_item.li_lsn);
2392
2393 xfs_buf_attach_iodone(bp, xfs_istale_done,
2394 &iip->ili_item);
2395
2396 if (ip != free_ip)
2397 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2398 }
2399
2400 xfs_trans_stale_inode_buf(tp, bp);
2401 xfs_trans_binval(tp, bp);
2402 }
2403
2404 xfs_perag_put(pag);
2405 return 0;
2406}
2407
2408/*
2409 * This is called to return an inode to the inode free list.
2410 * The inode should already be truncated to 0 length and have
2411 * no pages associated with it. This routine also assumes that
2412 * the inode is already a part of the transaction.
2413 *
2414 * The on-disk copy of the inode will have been added to the list
2415 * of unlinked inodes in the AGI. We need to remove the inode from
2416 * that list atomically with respect to freeing it here.
2417 */
2418int
2419xfs_ifree(
2420 xfs_trans_t *tp,
2421 xfs_inode_t *ip,
2422 xfs_bmap_free_t *flist)
2423{
2424 int error;
2425 int delete;
2426 xfs_ino_t first_ino;
2427
2428 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
2429 ASSERT(ip->i_d.di_nlink == 0);
2430 ASSERT(ip->i_d.di_nextents == 0);
2431 ASSERT(ip->i_d.di_anextents == 0);
2432 ASSERT(ip->i_d.di_size == 0 || !S_ISREG(ip->i_d.di_mode));
2433 ASSERT(ip->i_d.di_nblocks == 0);
2434
2435 /*
2436 * Pull the on-disk inode from the AGI unlinked list.
2437 */
2438 error = xfs_iunlink_remove(tp, ip);
2439 if (error)
2440 return error;
2441
2442 error = xfs_difree(tp, ip->i_ino, flist, &delete, &first_ino);
2443 if (error)
2444 return error;
2445
2446 ip->i_d.di_mode = 0; /* mark incore inode as free */
2447 ip->i_d.di_flags = 0;
2448 ip->i_d.di_dmevmask = 0;
2449 ip->i_d.di_forkoff = 0; /* mark the attr fork not in use */
2450 ip->i_d.di_format = XFS_DINODE_FMT_EXTENTS;
2451 ip->i_d.di_aformat = XFS_DINODE_FMT_EXTENTS;
2452 /*
2453 * Bump the generation count so no one will be confused
2454 * by reincarnations of this inode.
2455 */
2456 ip->i_d.di_gen++;
2457 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2458
2459 if (delete)
2460 error = xfs_ifree_cluster(ip, tp, first_ino);
2461
2462 return error;
2463}
2464
2465/*
2466 * This is called to unpin an inode. The caller must have the inode locked
2467 * in at least shared mode so that the buffer cannot be subsequently pinned
2468 * once someone is waiting for it to be unpinned.
2469 */
2470static void
2471xfs_iunpin(
2472 struct xfs_inode *ip)
2473{
2474 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
2475
2476 trace_xfs_inode_unpin_nowait(ip, _RET_IP_);
2477
2478 /* Give the log a push to start the unpinning I/O */
2479 xfs_log_force_lsn(ip->i_mount, ip->i_itemp->ili_last_lsn, 0);
2480
2481}
2482
2483static void
2484__xfs_iunpin_wait(
2485 struct xfs_inode *ip)
2486{
2487 wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_IPINNED_BIT);
2488 DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IPINNED_BIT);
2489
2490 xfs_iunpin(ip);
2491
2492 do {
2493 prepare_to_wait(wq, &wait.wait, TASK_UNINTERRUPTIBLE);
2494 if (xfs_ipincount(ip))
2495 io_schedule();
2496 } while (xfs_ipincount(ip));
2497 finish_wait(wq, &wait.wait);
2498}
2499
2500void
2501xfs_iunpin_wait(
2502 struct xfs_inode *ip)
2503{
2504 if (xfs_ipincount(ip))
2505 __xfs_iunpin_wait(ip);
2506}
2507
2508/*
2509 * Removing an inode from the namespace involves removing the directory entry
2510 * and dropping the link count on the inode. Removing the directory entry can
2511 * result in locking an AGF (directory blocks were freed) and removing a link
2512 * count can result in placing the inode on an unlinked list which results in
2513 * locking an AGI.
2514 *
2515 * The big problem here is that we have an ordering constraint on AGF and AGI
2516 * locking - inode allocation locks the AGI, then can allocate a new extent for
2517 * new inodes, locking the AGF after the AGI. Similarly, freeing the inode
2518 * removes the inode from the unlinked list, requiring that we lock the AGI
2519 * first, and then freeing the inode can result in an inode chunk being freed
2520 * and hence freeing disk space requiring that we lock an AGF.
2521 *
2522 * Hence the ordering that is imposed by other parts of the code is AGI before
2523 * AGF. This means we cannot remove the directory entry before we drop the inode
2524 * reference count and put it on the unlinked list as this results in a lock
2525 * order of AGF then AGI, and this can deadlock against inode allocation and
2526 * freeing. Therefore we must drop the link counts before we remove the
2527 * directory entry.
2528 *
2529 * This is still safe from a transactional point of view - it is not until we
2530 * get to xfs_bmap_finish() that we have the possibility of multiple
2531 * transactions in this operation. Hence as long as we remove the directory
2532 * entry and drop the link count in the first transaction of the remove
2533 * operation, there are no transactional constraints on the ordering here.
2534 */
2535int
2536xfs_remove(
2537 xfs_inode_t *dp,
2538 struct xfs_name *name,
2539 xfs_inode_t *ip)
2540{
2541 xfs_mount_t *mp = dp->i_mount;
2542 xfs_trans_t *tp = NULL;
2543 int is_dir = S_ISDIR(ip->i_d.di_mode);
2544 int error = 0;
2545 xfs_bmap_free_t free_list;
2546 xfs_fsblock_t first_block;
2547 int cancel_flags;
2548 int committed;
2549 int link_zero;
2550 uint resblks;
2551 uint log_count;
2552
2553 trace_xfs_remove(dp, name);
2554
2555 if (XFS_FORCED_SHUTDOWN(mp))
2556 return XFS_ERROR(EIO);
2557
2558 error = xfs_qm_dqattach(dp, 0);
2559 if (error)
2560 goto std_return;
2561
2562 error = xfs_qm_dqattach(ip, 0);
2563 if (error)
2564 goto std_return;
2565
2566 if (is_dir) {
2567 tp = xfs_trans_alloc(mp, XFS_TRANS_RMDIR);
2568 log_count = XFS_DEFAULT_LOG_COUNT;
2569 } else {
2570 tp = xfs_trans_alloc(mp, XFS_TRANS_REMOVE);
2571 log_count = XFS_REMOVE_LOG_COUNT;
2572 }
2573 cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
2574
2575 /*
2576 * We try to get the real space reservation first,
2577 * allowing for directory btree deletion(s) implying
2578 * possible bmap insert(s). If we can't get the space
2579 * reservation then we use 0 instead, and avoid the bmap
2580 * btree insert(s) in the directory code by, if the bmap
2581 * insert tries to happen, instead trimming the LAST
2582 * block from the directory.
2583 */
2584 resblks = XFS_REMOVE_SPACE_RES(mp);
2585 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_remove, resblks, 0);
2586 if (error == ENOSPC) {
2587 resblks = 0;
2588 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_remove, 0, 0);
2589 }
2590 if (error) {
2591 ASSERT(error != ENOSPC);
2592 cancel_flags = 0;
2593 goto out_trans_cancel;
2594 }
2595
2596 xfs_lock_two_inodes(dp, ip, XFS_ILOCK_EXCL);
2597
2598 xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
2599 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
2600
2601 /*
2602 * If we're removing a directory perform some additional validation.
2603 */
2604 cancel_flags |= XFS_TRANS_ABORT;
2605 if (is_dir) {
2606 ASSERT(ip->i_d.di_nlink >= 2);
2607 if (ip->i_d.di_nlink != 2) {
2608 error = XFS_ERROR(ENOTEMPTY);
2609 goto out_trans_cancel;
2610 }
2611 if (!xfs_dir_isempty(ip)) {
2612 error = XFS_ERROR(ENOTEMPTY);
2613 goto out_trans_cancel;
2614 }
2615
2616 /* Drop the link from ip's "..". */
2617 error = xfs_droplink(tp, dp);
2618 if (error)
2619 goto out_trans_cancel;
2620
2621 /* Drop the "." link from ip to self. */
2622 error = xfs_droplink(tp, ip);
2623 if (error)
2624 goto out_trans_cancel;
2625 } else {
2626 /*
2627 * When removing a non-directory we need to log the parent
2628 * inode here. For a directory this is done implicitly
2629 * by the xfs_droplink call for the ".." entry.
2630 */
2631 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
2632 }
2633 xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2634
2635 /* Drop the link from dp to ip. */
2636 error = xfs_droplink(tp, ip);
2637 if (error)
2638 goto out_trans_cancel;
2639
2640 /* Determine if this is the last link while the inode is locked */
2641 link_zero = (ip->i_d.di_nlink == 0);
2642
2643 xfs_bmap_init(&free_list, &first_block);
2644 error = xfs_dir_removename(tp, dp, name, ip->i_ino,
2645 &first_block, &free_list, resblks);
2646 if (error) {
2647 ASSERT(error != ENOENT);
2648 goto out_bmap_cancel;
2649 }
2650
2651 /*
2652 * If this is a synchronous mount, make sure that the
2653 * remove transaction goes to disk before returning to
2654 * the user.
2655 */
2656 if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC))
2657 xfs_trans_set_sync(tp);
2658
2659 error = xfs_bmap_finish(&tp, &free_list, &committed);
2660 if (error)
2661 goto out_bmap_cancel;
2662
2663 error = xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2664 if (error)
2665 goto std_return;
2666
2667 /*
2668 * If we are using filestreams, kill the stream association.
2669 * If the file is still open it may get a new one but that
2670 * will get killed on last close in xfs_close() so we don't
2671 * have to worry about that.
2672 */
2673 if (!is_dir && link_zero && xfs_inode_is_filestream(ip))
2674 xfs_filestream_deassociate(ip);
2675
2676 return 0;
2677
2678 out_bmap_cancel:
2679 xfs_bmap_cancel(&free_list);
2680 out_trans_cancel:
2681 xfs_trans_cancel(tp, cancel_flags);
2682 std_return:
2683 return error;
2684}
2685
2686/*
2687 * Enter all inodes for a rename transaction into a sorted array.
2688 */
2689STATIC void
2690xfs_sort_for_rename(
2691 xfs_inode_t *dp1, /* in: old (source) directory inode */
2692 xfs_inode_t *dp2, /* in: new (target) directory inode */
2693 xfs_inode_t *ip1, /* in: inode of old entry */
2694 xfs_inode_t *ip2, /* in: inode of new entry, if it
2695 already exists, NULL otherwise. */
2696 xfs_inode_t **i_tab,/* out: array of inode returned, sorted */
2697 int *num_inodes) /* out: number of inodes in array */
2698{
2699 xfs_inode_t *temp;
2700 int i, j;
2701
2702 /*
2703 * i_tab contains a list of pointers to inodes. We initialize
2704 * the table here & we'll sort it. We will then use it to
2705 * order the acquisition of the inode locks.
2706 *
2707 * Note that the table may contain duplicates. e.g., dp1 == dp2.
2708 */
2709 i_tab[0] = dp1;
2710 i_tab[1] = dp2;
2711 i_tab[2] = ip1;
2712 if (ip2) {
2713 *num_inodes = 4;
2714 i_tab[3] = ip2;
2715 } else {
2716 *num_inodes = 3;
2717 i_tab[3] = NULL;
2718 }
2719
2720 /*
2721 * Sort the elements via bubble sort. (Remember, there are at
2722 * most 4 elements to sort, so this is adequate.)
2723 */
2724 for (i = 0; i < *num_inodes; i++) {
2725 for (j = 1; j < *num_inodes; j++) {
2726 if (i_tab[j]->i_ino < i_tab[j-1]->i_ino) {
2727 temp = i_tab[j];
2728 i_tab[j] = i_tab[j-1];
2729 i_tab[j-1] = temp;
2730 }
2731 }
2732 }
2733}
2734
2735/*
2736 * xfs_rename
2737 */
2738int
2739xfs_rename(
2740 xfs_inode_t *src_dp,
2741 struct xfs_name *src_name,
2742 xfs_inode_t *src_ip,
2743 xfs_inode_t *target_dp,
2744 struct xfs_name *target_name,
2745 xfs_inode_t *target_ip)
2746{
2747 xfs_trans_t *tp = NULL;
2748 xfs_mount_t *mp = src_dp->i_mount;
2749 int new_parent; /* moving to a new dir */
2750 int src_is_directory; /* src_name is a directory */
2751 int error;
2752 xfs_bmap_free_t free_list;
2753 xfs_fsblock_t first_block;
2754 int cancel_flags;
2755 int committed;
2756 xfs_inode_t *inodes[4];
2757 int spaceres;
2758 int num_inodes;
2759
2760 trace_xfs_rename(src_dp, target_dp, src_name, target_name);
2761
2762 new_parent = (src_dp != target_dp);
2763 src_is_directory = S_ISDIR(src_ip->i_d.di_mode);
2764
2765 xfs_sort_for_rename(src_dp, target_dp, src_ip, target_ip,
2766 inodes, &num_inodes);
2767
2768 xfs_bmap_init(&free_list, &first_block);
2769 tp = xfs_trans_alloc(mp, XFS_TRANS_RENAME);
2770 cancel_flags = XFS_TRANS_RELEASE_LOG_RES;
2771 spaceres = XFS_RENAME_SPACE_RES(mp, target_name->len);
2772 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_rename, spaceres, 0);
2773 if (error == ENOSPC) {
2774 spaceres = 0;
2775 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_rename, 0, 0);
2776 }
2777 if (error) {
2778 xfs_trans_cancel(tp, 0);
2779 goto std_return;
2780 }
2781
2782 /*
2783 * Attach the dquots to the inodes
2784 */
2785 error = xfs_qm_vop_rename_dqattach(inodes);
2786 if (error) {
2787 xfs_trans_cancel(tp, cancel_flags);
2788 goto std_return;
2789 }
2790
2791 /*
2792 * Lock all the participating inodes. Depending upon whether
2793 * the target_name exists in the target directory, and
2794 * whether the target directory is the same as the source
2795 * directory, we can lock from 2 to 4 inodes.
2796 */
2797 xfs_lock_inodes(inodes, num_inodes, XFS_ILOCK_EXCL);
2798
2799 /*
2800 * Join all the inodes to the transaction. From this point on,
2801 * we can rely on either trans_commit or trans_cancel to unlock
2802 * them.
2803 */
2804 xfs_trans_ijoin(tp, src_dp, XFS_ILOCK_EXCL);
2805 if (new_parent)
2806 xfs_trans_ijoin(tp, target_dp, XFS_ILOCK_EXCL);
2807 xfs_trans_ijoin(tp, src_ip, XFS_ILOCK_EXCL);
2808 if (target_ip)
2809 xfs_trans_ijoin(tp, target_ip, XFS_ILOCK_EXCL);
2810
2811 /*
2812 * If we are using project inheritance, we only allow renames
2813 * into our tree when the project IDs are the same; else the
2814 * tree quota mechanism would be circumvented.
2815 */
2816 if (unlikely((target_dp->i_d.di_flags & XFS_DIFLAG_PROJINHERIT) &&
2817 (xfs_get_projid(target_dp) != xfs_get_projid(src_ip)))) {
2818 error = XFS_ERROR(EXDEV);
2819 goto error_return;
2820 }
2821
2822 /*
2823 * Set up the target.
2824 */
2825 if (target_ip == NULL) {
2826 /*
2827 * If there's no space reservation, check the entry will
2828 * fit before actually inserting it.
2829 */
2830 error = xfs_dir_canenter(tp, target_dp, target_name, spaceres);
2831 if (error)
2832 goto error_return;
2833 /*
2834 * If target does not exist and the rename crosses
2835 * directories, adjust the target directory link count
2836 * to account for the ".." reference from the new entry.
2837 */
2838 error = xfs_dir_createname(tp, target_dp, target_name,
2839 src_ip->i_ino, &first_block,
2840 &free_list, spaceres);
2841 if (error == ENOSPC)
2842 goto error_return;
2843 if (error)
2844 goto abort_return;
2845
2846 xfs_trans_ichgtime(tp, target_dp,
2847 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2848
2849 if (new_parent && src_is_directory) {
2850 error = xfs_bumplink(tp, target_dp);
2851 if (error)
2852 goto abort_return;
2853 }
2854 } else { /* target_ip != NULL */
2855 /*
2856 * If target exists and it's a directory, check that both
2857 * target and source are directories and that target can be
2858 * destroyed, or that neither is a directory.
2859 */
2860 if (S_ISDIR(target_ip->i_d.di_mode)) {
2861 /*
2862 * Make sure target dir is empty.
2863 */
2864 if (!(xfs_dir_isempty(target_ip)) ||
2865 (target_ip->i_d.di_nlink > 2)) {
2866 error = XFS_ERROR(EEXIST);
2867 goto error_return;
2868 }
2869 }
2870
2871 /*
2872 * Link the source inode under the target name.
2873 * If the source inode is a directory and we are moving
2874 * it across directories, its ".." entry will be
2875 * inconsistent until we replace that down below.
2876 *
2877 * In case there is already an entry with the same
2878 * name at the destination directory, remove it first.
2879 */
2880 error = xfs_dir_replace(tp, target_dp, target_name,
2881 src_ip->i_ino,
2882 &first_block, &free_list, spaceres);
2883 if (error)
2884 goto abort_return;
2885
2886 xfs_trans_ichgtime(tp, target_dp,
2887 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2888
2889 /*
2890 * Decrement the link count on the target since the target
2891 * dir no longer points to it.
2892 */
2893 error = xfs_droplink(tp, target_ip);
2894 if (error)
2895 goto abort_return;
2896
2897 if (src_is_directory) {
2898 /*
2899 * Drop the link from the old "." entry.
2900 */
2901 error = xfs_droplink(tp, target_ip);
2902 if (error)
2903 goto abort_return;
2904 }
2905 } /* target_ip != NULL */
2906
2907 /*
2908 * Remove the source.
2909 */
2910 if (new_parent && src_is_directory) {
2911 /*
2912 * Rewrite the ".." entry to point to the new
2913 * directory.
2914 */
2915 error = xfs_dir_replace(tp, src_ip, &xfs_name_dotdot,
2916 target_dp->i_ino,
2917 &first_block, &free_list, spaceres);
2918 ASSERT(error != EEXIST);
2919 if (error)
2920 goto abort_return;
2921 }
2922
2923 /*
2924 * We always want to hit the ctime on the source inode.
2925 *
2926 * This isn't strictly required by the standards since the source
2927 * inode isn't really being changed, but old unix file systems did
2928 * it and some incremental backup programs won't work without it.
2929 */
2930 xfs_trans_ichgtime(tp, src_ip, XFS_ICHGTIME_CHG);
2931 xfs_trans_log_inode(tp, src_ip, XFS_ILOG_CORE);
2932
2933 /*
2934 * Adjust the link count on src_dp. This is necessary when
2935 * renaming a directory, either within one parent when
2936 * the target existed, or across two parent directories.
2937 */
2938 if (src_is_directory && (new_parent || target_ip != NULL)) {
2939
2940 /*
2941 * Decrement link count on src_directory since the
2942 * entry that's moved no longer points to it.
2943 */
2944 error = xfs_droplink(tp, src_dp);
2945 if (error)
2946 goto abort_return;
2947 }
2948
2949 error = xfs_dir_removename(tp, src_dp, src_name, src_ip->i_ino,
2950 &first_block, &free_list, spaceres);
2951 if (error)
2952 goto abort_return;
2953
2954 xfs_trans_ichgtime(tp, src_dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2955 xfs_trans_log_inode(tp, src_dp, XFS_ILOG_CORE);
2956 if (new_parent)
2957 xfs_trans_log_inode(tp, target_dp, XFS_ILOG_CORE);
2958
2959 /*
2960 * If this is a synchronous mount, make sure that the
2961 * rename transaction goes to disk before returning to
2962 * the user.
2963 */
2964 if (mp->m_flags & (XFS_MOUNT_WSYNC|XFS_MOUNT_DIRSYNC)) {
2965 xfs_trans_set_sync(tp);
2966 }
2967
2968 error = xfs_bmap_finish(&tp, &free_list, &committed);
2969 if (error) {
2970 xfs_bmap_cancel(&free_list);
2971 xfs_trans_cancel(tp, (XFS_TRANS_RELEASE_LOG_RES |
2972 XFS_TRANS_ABORT));
2973 goto std_return;
2974 }
2975
2976 /*
2977 * trans_commit will unlock src_ip, target_ip & decrement
2978 * the vnode references.
2979 */
2980 return xfs_trans_commit(tp, XFS_TRANS_RELEASE_LOG_RES);
2981
2982 abort_return:
2983 cancel_flags |= XFS_TRANS_ABORT;
2984 error_return:
2985 xfs_bmap_cancel(&free_list);
2986 xfs_trans_cancel(tp, cancel_flags);
2987 std_return:
2988 return error;
2989}
2990
2991STATIC int
2992xfs_iflush_cluster(
2993 xfs_inode_t *ip,
2994 xfs_buf_t *bp)
2995{
2996 xfs_mount_t *mp = ip->i_mount;
2997 struct xfs_perag *pag;
2998 unsigned long first_index, mask;
2999 unsigned long inodes_per_cluster;
3000 int ilist_size;
3001 xfs_inode_t **ilist;
3002 xfs_inode_t *iq;
3003 int nr_found;
3004 int clcount = 0;
3005 int bufwasdelwri;
3006 int i;
3007
3008 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
3009
3010 inodes_per_cluster = mp->m_inode_cluster_size >> mp->m_sb.sb_inodelog;
3011 ilist_size = inodes_per_cluster * sizeof(xfs_inode_t *);
3012 ilist = kmem_alloc(ilist_size, KM_MAYFAIL|KM_NOFS);
3013 if (!ilist)
3014 goto out_put;
3015
3016 mask = ~(((mp->m_inode_cluster_size >> mp->m_sb.sb_inodelog)) - 1);
3017 first_index = XFS_INO_TO_AGINO(mp, ip->i_ino) & mask;
3018 rcu_read_lock();
3019 /* really need a gang lookup range call here */
3020 nr_found = radix_tree_gang_lookup(&pag->pag_ici_root, (void**)ilist,
3021 first_index, inodes_per_cluster);
3022 if (nr_found == 0)
3023 goto out_free;
3024
3025 for (i = 0; i < nr_found; i++) {
3026 iq = ilist[i];
3027 if (iq == ip)
3028 continue;
3029
3030 /*
3031 * because this is an RCU protected lookup, we could find a
3032 * recently freed or even reallocated inode during the lookup.
3033 * We need to check under the i_flags_lock for a valid inode
3034 * here. Skip it if it is not valid or the wrong inode.
3035 */
3036 spin_lock(&ip->i_flags_lock);
3037 if (!ip->i_ino ||
3038 (XFS_INO_TO_AGINO(mp, iq->i_ino) & mask) != first_index) {
3039 spin_unlock(&ip->i_flags_lock);
3040 continue;
3041 }
3042 spin_unlock(&ip->i_flags_lock);
3043
3044 /*
3045 * Do an un-protected check to see if the inode is dirty and
3046 * is a candidate for flushing. These checks will be repeated
3047 * later after the appropriate locks are acquired.
3048 */
3049 if (xfs_inode_clean(iq) && xfs_ipincount(iq) == 0)
3050 continue;
3051
3052 /*
3053 * Try to get locks. If any are unavailable or it is pinned,
3054 * then this inode cannot be flushed and is skipped.
3055 */
3056
3057 if (!xfs_ilock_nowait(iq, XFS_ILOCK_SHARED))
3058 continue;
3059 if (!xfs_iflock_nowait(iq)) {
3060 xfs_iunlock(iq, XFS_ILOCK_SHARED);
3061 continue;
3062 }
3063 if (xfs_ipincount(iq)) {
3064 xfs_ifunlock(iq);
3065 xfs_iunlock(iq, XFS_ILOCK_SHARED);
3066 continue;
3067 }
3068
3069 /*
3070 * arriving here means that this inode can be flushed. First
3071 * re-check that it's dirty before flushing.
3072 */
3073 if (!xfs_inode_clean(iq)) {
3074 int error;
3075 error = xfs_iflush_int(iq, bp);
3076 if (error) {
3077 xfs_iunlock(iq, XFS_ILOCK_SHARED);
3078 goto cluster_corrupt_out;
3079 }
3080 clcount++;
3081 } else {
3082 xfs_ifunlock(iq);
3083 }
3084 xfs_iunlock(iq, XFS_ILOCK_SHARED);
3085 }
3086
3087 if (clcount) {
3088 XFS_STATS_INC(xs_icluster_flushcnt);
3089 XFS_STATS_ADD(xs_icluster_flushinode, clcount);
3090 }
3091
3092out_free:
3093 rcu_read_unlock();
3094 kmem_free(ilist);
3095out_put:
3096 xfs_perag_put(pag);
3097 return 0;
3098
3099
3100cluster_corrupt_out:
3101 /*
3102 * Corruption detected in the clustering loop. Invalidate the
3103 * inode buffer and shut down the filesystem.
3104 */
3105 rcu_read_unlock();
3106 /*
3107 * Clean up the buffer. If it was delwri, just release it --
3108 * brelse can handle it with no problems. If not, shut down the
3109 * filesystem before releasing the buffer.
3110 */
3111 bufwasdelwri = (bp->b_flags & _XBF_DELWRI_Q);
3112 if (bufwasdelwri)
3113 xfs_buf_relse(bp);
3114
3115 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
3116
3117 if (!bufwasdelwri) {
3118 /*
3119 * Just like incore_relse: if we have b_iodone functions,
3120 * mark the buffer as an error and call them. Otherwise
3121 * mark it as stale and brelse.
3122 */
3123 if (bp->b_iodone) {
3124 XFS_BUF_UNDONE(bp);
3125 xfs_buf_stale(bp);
3126 xfs_buf_ioerror(bp, EIO);
3127 xfs_buf_ioend(bp, 0);
3128 } else {
3129 xfs_buf_stale(bp);
3130 xfs_buf_relse(bp);
3131 }
3132 }
3133
3134 /*
3135 * Unlocks the flush lock
3136 */
3137 xfs_iflush_abort(iq, false);
3138 kmem_free(ilist);
3139 xfs_perag_put(pag);
3140 return XFS_ERROR(EFSCORRUPTED);
3141}
3142
3143/*
3144 * Flush dirty inode metadata into the backing buffer.
3145 *
3146 * The caller must have the inode lock and the inode flush lock held. The
3147 * inode lock will still be held upon return to the caller, and the inode
3148 * flush lock will be released after the inode has reached the disk.
3149 *
3150 * The caller must write out the buffer returned in *bpp and release it.
3151 */
3152int
3153xfs_iflush(
3154 struct xfs_inode *ip,
3155 struct xfs_buf **bpp)
3156{
3157 struct xfs_mount *mp = ip->i_mount;
3158 struct xfs_buf *bp;
3159 struct xfs_dinode *dip;
3160 int error;
3161
3162 XFS_STATS_INC(xs_iflush_count);
3163
3164 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
3165 ASSERT(xfs_isiflocked(ip));
3166 ASSERT(ip->i_d.di_format != XFS_DINODE_FMT_BTREE ||
3167 ip->i_d.di_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
3168
3169 *bpp = NULL;
3170
3171 xfs_iunpin_wait(ip);
3172
3173 /*
3174 * For stale inodes we cannot rely on the backing buffer remaining
3175 * stale in cache for the remaining life of the stale inode and so
3176 * xfs_imap_to_bp() below may give us a buffer that no longer contains
3177 * inodes below. We have to check this after ensuring the inode is
3178 * unpinned so that it is safe to reclaim the stale inode after the
3179 * flush call.
3180 */
3181 if (xfs_iflags_test(ip, XFS_ISTALE)) {
3182 xfs_ifunlock(ip);
3183 return 0;
3184 }
3185
3186 /*
3187 * This may have been unpinned because the filesystem is shutting
3188 * down forcibly. If that's the case we must not write this inode
3189 * to disk, because the log record didn't make it to disk.
3190 *
3191 * We also have to remove the log item from the AIL in this case,
3192 * as we wait for an empty AIL as part of the unmount process.
3193 */
3194 if (XFS_FORCED_SHUTDOWN(mp)) {
3195 error = XFS_ERROR(EIO);
3196 goto abort_out;
3197 }
3198
3199 /*
3200 * Get the buffer containing the on-disk inode.
3201 */
3202 error = xfs_imap_to_bp(mp, NULL, &ip->i_imap, &dip, &bp, XBF_TRYLOCK,
3203 0);
3204 if (error || !bp) {
3205 xfs_ifunlock(ip);
3206 return error;
3207 }
3208
3209 /*
3210 * First flush out the inode that xfs_iflush was called with.
3211 */
3212 error = xfs_iflush_int(ip, bp);
3213 if (error)
3214 goto corrupt_out;
3215
3216 /*
3217 * If the buffer is pinned then push on the log now so we won't
3218 * get stuck waiting in the write for too long.
3219 */
3220 if (xfs_buf_ispinned(bp))
3221 xfs_log_force(mp, 0);
3222
3223 /*
3224 * inode clustering:
3225 * see if other inodes can be gathered into this write
3226 */
3227 error = xfs_iflush_cluster(ip, bp);
3228 if (error)
3229 goto cluster_corrupt_out;
3230
3231 *bpp = bp;
3232 return 0;
3233
3234corrupt_out:
3235 xfs_buf_relse(bp);
3236 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
3237cluster_corrupt_out:
3238 error = XFS_ERROR(EFSCORRUPTED);
3239abort_out:
3240 /*
3241 * Unlocks the flush lock
3242 */
3243 xfs_iflush_abort(ip, false);
3244 return error;
3245}
3246
3247STATIC int
3248xfs_iflush_int(
3249 struct xfs_inode *ip,
3250 struct xfs_buf *bp)
3251{
3252 struct xfs_inode_log_item *iip = ip->i_itemp;
3253 struct xfs_dinode *dip;
3254 struct xfs_mount *mp = ip->i_mount;
3255
3256 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
3257 ASSERT(xfs_isiflocked(ip));
3258 ASSERT(ip->i_d.di_format != XFS_DINODE_FMT_BTREE ||
3259 ip->i_d.di_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
3260 ASSERT(iip != NULL && iip->ili_fields != 0);
3261
3262 /* set *dip = inode's place in the buffer */
3263 dip = (xfs_dinode_t *)xfs_buf_offset(bp, ip->i_imap.im_boffset);
3264
3265 if (XFS_TEST_ERROR(dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC),
3266 mp, XFS_ERRTAG_IFLUSH_1, XFS_RANDOM_IFLUSH_1)) {
3267 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3268 "%s: Bad inode %Lu magic number 0x%x, ptr 0x%p",
3269 __func__, ip->i_ino, be16_to_cpu(dip->di_magic), dip);
3270 goto corrupt_out;
3271 }
3272 if (XFS_TEST_ERROR(ip->i_d.di_magic != XFS_DINODE_MAGIC,
3273 mp, XFS_ERRTAG_IFLUSH_2, XFS_RANDOM_IFLUSH_2)) {
3274 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3275 "%s: Bad inode %Lu, ptr 0x%p, magic number 0x%x",
3276 __func__, ip->i_ino, ip, ip->i_d.di_magic);
3277 goto corrupt_out;
3278 }
3279 if (S_ISREG(ip->i_d.di_mode)) {
3280 if (XFS_TEST_ERROR(
3281 (ip->i_d.di_format != XFS_DINODE_FMT_EXTENTS) &&
3282 (ip->i_d.di_format != XFS_DINODE_FMT_BTREE),
3283 mp, XFS_ERRTAG_IFLUSH_3, XFS_RANDOM_IFLUSH_3)) {
3284 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3285 "%s: Bad regular inode %Lu, ptr 0x%p",
3286 __func__, ip->i_ino, ip);
3287 goto corrupt_out;
3288 }
3289 } else if (S_ISDIR(ip->i_d.di_mode)) {
3290 if (XFS_TEST_ERROR(
3291 (ip->i_d.di_format != XFS_DINODE_FMT_EXTENTS) &&
3292 (ip->i_d.di_format != XFS_DINODE_FMT_BTREE) &&
3293 (ip->i_d.di_format != XFS_DINODE_FMT_LOCAL),
3294 mp, XFS_ERRTAG_IFLUSH_4, XFS_RANDOM_IFLUSH_4)) {
3295 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3296 "%s: Bad directory inode %Lu, ptr 0x%p",
3297 __func__, ip->i_ino, ip);
3298 goto corrupt_out;
3299 }
3300 }
3301 if (XFS_TEST_ERROR(ip->i_d.di_nextents + ip->i_d.di_anextents >
3302 ip->i_d.di_nblocks, mp, XFS_ERRTAG_IFLUSH_5,
3303 XFS_RANDOM_IFLUSH_5)) {
3304 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3305 "%s: detected corrupt incore inode %Lu, "
3306 "total extents = %d, nblocks = %Ld, ptr 0x%p",
3307 __func__, ip->i_ino,
3308 ip->i_d.di_nextents + ip->i_d.di_anextents,
3309 ip->i_d.di_nblocks, ip);
3310 goto corrupt_out;
3311 }
3312 if (XFS_TEST_ERROR(ip->i_d.di_forkoff > mp->m_sb.sb_inodesize,
3313 mp, XFS_ERRTAG_IFLUSH_6, XFS_RANDOM_IFLUSH_6)) {
3314 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3315 "%s: bad inode %Lu, forkoff 0x%x, ptr 0x%p",
3316 __func__, ip->i_ino, ip->i_d.di_forkoff, ip);
3317 goto corrupt_out;
3318 }
3319
3320 /*
3321 * Inode item log recovery for v1/v2 inodes are dependent on the
3322 * di_flushiter count for correct sequencing. We bump the flush
3323 * iteration count so we can detect flushes which postdate a log record
3324 * during recovery. This is redundant as we now log every change and
3325 * hence this can't happen but we need to still do it to ensure
3326 * backwards compatibility with old kernels that predate logging all
3327 * inode changes.
3328 */
3329 if (ip->i_d.di_version < 3)
3330 ip->i_d.di_flushiter++;
3331
3332 /*
3333 * Copy the dirty parts of the inode into the on-disk
3334 * inode. We always copy out the core of the inode,
3335 * because if the inode is dirty at all the core must
3336 * be.
3337 */
3338 xfs_dinode_to_disk(dip, &ip->i_d);
3339
3340 /* Wrap, we never let the log put out DI_MAX_FLUSH */
3341 if (ip->i_d.di_flushiter == DI_MAX_FLUSH)
3342 ip->i_d.di_flushiter = 0;
3343
3344 /*
3345 * If this is really an old format inode and the superblock version
3346 * has not been updated to support only new format inodes, then
3347 * convert back to the old inode format. If the superblock version
3348 * has been updated, then make the conversion permanent.
3349 */
3350 ASSERT(ip->i_d.di_version == 1 || xfs_sb_version_hasnlink(&mp->m_sb));
3351 if (ip->i_d.di_version == 1) {
3352 if (!xfs_sb_version_hasnlink(&mp->m_sb)) {
3353 /*
3354 * Convert it back.
3355 */
3356 ASSERT(ip->i_d.di_nlink <= XFS_MAXLINK_1);
3357 dip->di_onlink = cpu_to_be16(ip->i_d.di_nlink);
3358 } else {
3359 /*
3360 * The superblock version has already been bumped,
3361 * so just make the conversion to the new inode
3362 * format permanent.
3363 */
3364 ip->i_d.di_version = 2;
3365 dip->di_version = 2;
3366 ip->i_d.di_onlink = 0;
3367 dip->di_onlink = 0;
3368 memset(&(ip->i_d.di_pad[0]), 0, sizeof(ip->i_d.di_pad));
3369 memset(&(dip->di_pad[0]), 0,
3370 sizeof(dip->di_pad));
3371 ASSERT(xfs_get_projid(ip) == 0);
3372 }
3373 }
3374
3375 xfs_iflush_fork(ip, dip, iip, XFS_DATA_FORK, bp);
3376 if (XFS_IFORK_Q(ip))
3377 xfs_iflush_fork(ip, dip, iip, XFS_ATTR_FORK, bp);
3378 xfs_inobp_check(mp, bp);
3379
3380 /*
3381 * We've recorded everything logged in the inode, so we'd like to clear
3382 * the ili_fields bits so we don't log and flush things unnecessarily.
3383 * However, we can't stop logging all this information until the data
3384 * we've copied into the disk buffer is written to disk. If we did we
3385 * might overwrite the copy of the inode in the log with all the data
3386 * after re-logging only part of it, and in the face of a crash we
3387 * wouldn't have all the data we need to recover.
3388 *
3389 * What we do is move the bits to the ili_last_fields field. When
3390 * logging the inode, these bits are moved back to the ili_fields field.
3391 * In the xfs_iflush_done() routine we clear ili_last_fields, since we
3392 * know that the information those bits represent is permanently on
3393 * disk. As long as the flush completes before the inode is logged
3394 * again, then both ili_fields and ili_last_fields will be cleared.
3395 *
3396 * We can play with the ili_fields bits here, because the inode lock
3397 * must be held exclusively in order to set bits there and the flush
3398 * lock protects the ili_last_fields bits. Set ili_logged so the flush
3399 * done routine can tell whether or not to look in the AIL. Also, store
3400 * the current LSN of the inode so that we can tell whether the item has
3401 * moved in the AIL from xfs_iflush_done(). In order to read the lsn we
3402 * need the AIL lock, because it is a 64 bit value that cannot be read
3403 * atomically.
3404 */
3405 iip->ili_last_fields = iip->ili_fields;
3406 iip->ili_fields = 0;
3407 iip->ili_logged = 1;
3408
3409 xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
3410 &iip->ili_item.li_lsn);
3411
3412 /*
3413 * Attach the function xfs_iflush_done to the inode's
3414 * buffer. This will remove the inode from the AIL
3415 * and unlock the inode's flush lock when the inode is
3416 * completely written to disk.
3417 */
3418 xfs_buf_attach_iodone(bp, xfs_iflush_done, &iip->ili_item);
3419
3420 /* update the lsn in the on disk inode if required */
3421 if (ip->i_d.di_version == 3)
3422 dip->di_lsn = cpu_to_be64(iip->ili_item.li_lsn);
3423
3424 /* generate the checksum. */
3425 xfs_dinode_calc_crc(mp, dip);
3426
3427 ASSERT(bp->b_fspriv != NULL);
3428 ASSERT(bp->b_iodone != NULL);
3429 return 0;
3430
3431corrupt_out:
3432 return XFS_ERROR(EFSCORRUPTED);
3433}
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}