1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Copyright (c) 2000-2006 Silicon Graphics, Inc.
4 * All Rights Reserved.
5 */
6 #include <linux/iversion.h>
7
8 #include "xfs.h"
9 #include "xfs_fs.h"
10 #include "xfs_shared.h"
11 #include "xfs_format.h"
12 #include "xfs_log_format.h"
13 #include "xfs_trans_resv.h"
14 #include "xfs_mount.h"
15 #include "xfs_defer.h"
16 #include "xfs_inode.h"
17 #include "xfs_dir2.h"
18 #include "xfs_attr.h"
19 #include "xfs_trans_space.h"
20 #include "xfs_trans.h"
21 #include "xfs_buf_item.h"
22 #include "xfs_inode_item.h"
23 #include "xfs_ialloc.h"
24 #include "xfs_bmap.h"
25 #include "xfs_bmap_util.h"
26 #include "xfs_errortag.h"
27 #include "xfs_error.h"
28 #include "xfs_quota.h"
29 #include "xfs_filestream.h"
30 #include "xfs_trace.h"
31 #include "xfs_icache.h"
32 #include "xfs_symlink.h"
33 #include "xfs_trans_priv.h"
34 #include "xfs_log.h"
35 #include "xfs_bmap_btree.h"
36 #include "xfs_reflink.h"
37 #include "xfs_ag.h"
38 #include "xfs_log_priv.h"
39
40 struct kmem_cache *xfs_inode_cache;
41
42 /*
43 * Used in xfs_itruncate_extents(). This is the maximum number of extents
44 * freed from a file in a single transaction.
45 */
46 #define XFS_ITRUNC_MAX_EXTENTS 2
47
48 STATIC int xfs_iunlink(struct xfs_trans *, struct xfs_inode *);
49 STATIC int xfs_iunlink_remove(struct xfs_trans *tp, struct xfs_perag *pag,
50 struct xfs_inode *);
51
52 /*
53 * helper function to extract extent size hint from inode
54 */
55 xfs_extlen_t
xfs_get_extsz_hint(struct xfs_inode * ip)56 xfs_get_extsz_hint(
57 struct xfs_inode *ip)
58 {
59 /*
60 * No point in aligning allocations if we need to COW to actually
61 * write to them.
62 */
63 if (xfs_is_always_cow_inode(ip))
64 return 0;
65 if ((ip->i_diflags & XFS_DIFLAG_EXTSIZE) && ip->i_extsize)
66 return ip->i_extsize;
67 if (XFS_IS_REALTIME_INODE(ip))
68 return ip->i_mount->m_sb.sb_rextsize;
69 return 0;
70 }
71
72 /*
73 * Helper function to extract CoW extent size hint from inode.
74 * Between the extent size hint and the CoW extent size hint, we
75 * return the greater of the two. If the value is zero (automatic),
76 * use the default size.
77 */
78 xfs_extlen_t
xfs_get_cowextsz_hint(struct xfs_inode * ip)79 xfs_get_cowextsz_hint(
80 struct xfs_inode *ip)
81 {
82 xfs_extlen_t a, b;
83
84 a = 0;
85 if (ip->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE)
86 a = ip->i_cowextsize;
87 b = xfs_get_extsz_hint(ip);
88
89 a = max(a, b);
90 if (a == 0)
91 return XFS_DEFAULT_COWEXTSZ_HINT;
92 return a;
93 }
94
95 /*
96 * These two are wrapper routines around the xfs_ilock() routine used to
97 * centralize some grungy code. They are used in places that wish to lock the
98 * inode solely for reading the extents. The reason these places can't just
99 * call xfs_ilock(ip, XFS_ILOCK_SHARED) is that the inode lock also guards to
100 * bringing in of the extents from disk for a file in b-tree format. If the
101 * inode is in b-tree format, then we need to lock the inode exclusively until
102 * the extents are read in. Locking it exclusively all the time would limit
103 * our parallelism unnecessarily, though. What we do instead is check to see
104 * if the extents have been read in yet, and only lock the inode exclusively
105 * if they have not.
106 *
107 * The functions return a value which should be given to the corresponding
108 * xfs_iunlock() call.
109 */
110 uint
xfs_ilock_data_map_shared(struct xfs_inode * ip)111 xfs_ilock_data_map_shared(
112 struct xfs_inode *ip)
113 {
114 uint lock_mode = XFS_ILOCK_SHARED;
115
116 if (xfs_need_iread_extents(&ip->i_df))
117 lock_mode = XFS_ILOCK_EXCL;
118 xfs_ilock(ip, lock_mode);
119 return lock_mode;
120 }
121
122 uint
xfs_ilock_attr_map_shared(struct xfs_inode * ip)123 xfs_ilock_attr_map_shared(
124 struct xfs_inode *ip)
125 {
126 uint lock_mode = XFS_ILOCK_SHARED;
127
128 if (ip->i_afp && xfs_need_iread_extents(ip->i_afp))
129 lock_mode = XFS_ILOCK_EXCL;
130 xfs_ilock(ip, lock_mode);
131 return lock_mode;
132 }
133
134 /*
135 * You can't set both SHARED and EXCL for the same lock,
136 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_MMAPLOCK_SHARED,
137 * XFS_MMAPLOCK_EXCL, XFS_ILOCK_SHARED, XFS_ILOCK_EXCL are valid values
138 * to set in lock_flags.
139 */
140 static inline void
xfs_lock_flags_assert(uint lock_flags)141 xfs_lock_flags_assert(
142 uint lock_flags)
143 {
144 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
145 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
146 ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
147 (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
148 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
149 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
150 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
151 ASSERT(lock_flags != 0);
152 }
153
154 /*
155 * In addition to i_rwsem in the VFS inode, the xfs inode contains 2
156 * multi-reader locks: invalidate_lock and the i_lock. This routine allows
157 * various combinations of the locks to be obtained.
158 *
159 * The 3 locks should always be ordered so that the IO lock is obtained first,
160 * the mmap lock second and the ilock last in order to prevent deadlock.
161 *
162 * Basic locking order:
163 *
164 * i_rwsem -> invalidate_lock -> page_lock -> i_ilock
165 *
166 * mmap_lock locking order:
167 *
168 * i_rwsem -> page lock -> mmap_lock
169 * mmap_lock -> invalidate_lock -> page_lock
170 *
171 * The difference in mmap_lock locking order mean that we cannot hold the
172 * invalidate_lock over syscall based read(2)/write(2) based IO. These IO paths
173 * can fault in pages during copy in/out (for buffered IO) or require the
174 * mmap_lock in get_user_pages() to map the user pages into the kernel address
175 * space for direct IO. Similarly the i_rwsem cannot be taken inside a page
176 * fault because page faults already hold the mmap_lock.
177 *
178 * Hence to serialise fully against both syscall and mmap based IO, we need to
179 * take both the i_rwsem and the invalidate_lock. These locks should *only* be
180 * both taken in places where we need to invalidate the page cache in a race
181 * free manner (e.g. truncate, hole punch and other extent manipulation
182 * functions).
183 */
184 void
xfs_ilock(xfs_inode_t * ip,uint lock_flags)185 xfs_ilock(
186 xfs_inode_t *ip,
187 uint lock_flags)
188 {
189 trace_xfs_ilock(ip, lock_flags, _RET_IP_);
190
191 xfs_lock_flags_assert(lock_flags);
192
193 if (lock_flags & XFS_IOLOCK_EXCL) {
194 down_write_nested(&VFS_I(ip)->i_rwsem,
195 XFS_IOLOCK_DEP(lock_flags));
196 } else if (lock_flags & XFS_IOLOCK_SHARED) {
197 down_read_nested(&VFS_I(ip)->i_rwsem,
198 XFS_IOLOCK_DEP(lock_flags));
199 }
200
201 if (lock_flags & XFS_MMAPLOCK_EXCL) {
202 down_write_nested(&VFS_I(ip)->i_mapping->invalidate_lock,
203 XFS_MMAPLOCK_DEP(lock_flags));
204 } else if (lock_flags & XFS_MMAPLOCK_SHARED) {
205 down_read_nested(&VFS_I(ip)->i_mapping->invalidate_lock,
206 XFS_MMAPLOCK_DEP(lock_flags));
207 }
208
209 if (lock_flags & XFS_ILOCK_EXCL)
210 mrupdate_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
211 else if (lock_flags & XFS_ILOCK_SHARED)
212 mraccess_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
213 }
214
215 /*
216 * This is just like xfs_ilock(), except that the caller
217 * is guaranteed not to sleep. It returns 1 if it gets
218 * the requested locks and 0 otherwise. If the IO lock is
219 * obtained but the inode lock cannot be, then the IO lock
220 * is dropped before returning.
221 *
222 * ip -- the inode being locked
223 * lock_flags -- this parameter indicates the inode's locks to be
224 * to be locked. See the comment for xfs_ilock() for a list
225 * of valid values.
226 */
227 int
xfs_ilock_nowait(xfs_inode_t * ip,uint lock_flags)228 xfs_ilock_nowait(
229 xfs_inode_t *ip,
230 uint lock_flags)
231 {
232 trace_xfs_ilock_nowait(ip, lock_flags, _RET_IP_);
233
234 xfs_lock_flags_assert(lock_flags);
235
236 if (lock_flags & XFS_IOLOCK_EXCL) {
237 if (!down_write_trylock(&VFS_I(ip)->i_rwsem))
238 goto out;
239 } else if (lock_flags & XFS_IOLOCK_SHARED) {
240 if (!down_read_trylock(&VFS_I(ip)->i_rwsem))
241 goto out;
242 }
243
244 if (lock_flags & XFS_MMAPLOCK_EXCL) {
245 if (!down_write_trylock(&VFS_I(ip)->i_mapping->invalidate_lock))
246 goto out_undo_iolock;
247 } else if (lock_flags & XFS_MMAPLOCK_SHARED) {
248 if (!down_read_trylock(&VFS_I(ip)->i_mapping->invalidate_lock))
249 goto out_undo_iolock;
250 }
251
252 if (lock_flags & XFS_ILOCK_EXCL) {
253 if (!mrtryupdate(&ip->i_lock))
254 goto out_undo_mmaplock;
255 } else if (lock_flags & XFS_ILOCK_SHARED) {
256 if (!mrtryaccess(&ip->i_lock))
257 goto out_undo_mmaplock;
258 }
259 return 1;
260
261 out_undo_mmaplock:
262 if (lock_flags & XFS_MMAPLOCK_EXCL)
263 up_write(&VFS_I(ip)->i_mapping->invalidate_lock);
264 else if (lock_flags & XFS_MMAPLOCK_SHARED)
265 up_read(&VFS_I(ip)->i_mapping->invalidate_lock);
266 out_undo_iolock:
267 if (lock_flags & XFS_IOLOCK_EXCL)
268 up_write(&VFS_I(ip)->i_rwsem);
269 else if (lock_flags & XFS_IOLOCK_SHARED)
270 up_read(&VFS_I(ip)->i_rwsem);
271 out:
272 return 0;
273 }
274
275 /*
276 * xfs_iunlock() is used to drop the inode locks acquired with
277 * xfs_ilock() and xfs_ilock_nowait(). The caller must pass
278 * in the flags given to xfs_ilock() or xfs_ilock_nowait() so
279 * that we know which locks to drop.
280 *
281 * ip -- the inode being unlocked
282 * lock_flags -- this parameter indicates the inode's locks to be
283 * to be unlocked. See the comment for xfs_ilock() for a list
284 * of valid values for this parameter.
285 *
286 */
287 void
xfs_iunlock(xfs_inode_t * ip,uint lock_flags)288 xfs_iunlock(
289 xfs_inode_t *ip,
290 uint lock_flags)
291 {
292 xfs_lock_flags_assert(lock_flags);
293
294 if (lock_flags & XFS_IOLOCK_EXCL)
295 up_write(&VFS_I(ip)->i_rwsem);
296 else if (lock_flags & XFS_IOLOCK_SHARED)
297 up_read(&VFS_I(ip)->i_rwsem);
298
299 if (lock_flags & XFS_MMAPLOCK_EXCL)
300 up_write(&VFS_I(ip)->i_mapping->invalidate_lock);
301 else if (lock_flags & XFS_MMAPLOCK_SHARED)
302 up_read(&VFS_I(ip)->i_mapping->invalidate_lock);
303
304 if (lock_flags & XFS_ILOCK_EXCL)
305 mrunlock_excl(&ip->i_lock);
306 else if (lock_flags & XFS_ILOCK_SHARED)
307 mrunlock_shared(&ip->i_lock);
308
309 trace_xfs_iunlock(ip, lock_flags, _RET_IP_);
310 }
311
312 /*
313 * give up write locks. the i/o lock cannot be held nested
314 * if it is being demoted.
315 */
316 void
xfs_ilock_demote(xfs_inode_t * ip,uint lock_flags)317 xfs_ilock_demote(
318 xfs_inode_t *ip,
319 uint lock_flags)
320 {
321 ASSERT(lock_flags & (XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL));
322 ASSERT((lock_flags &
323 ~(XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL)) == 0);
324
325 if (lock_flags & XFS_ILOCK_EXCL)
326 mrdemote(&ip->i_lock);
327 if (lock_flags & XFS_MMAPLOCK_EXCL)
328 downgrade_write(&VFS_I(ip)->i_mapping->invalidate_lock);
329 if (lock_flags & XFS_IOLOCK_EXCL)
330 downgrade_write(&VFS_I(ip)->i_rwsem);
331
332 trace_xfs_ilock_demote(ip, lock_flags, _RET_IP_);
333 }
334
335 #if defined(DEBUG) || defined(XFS_WARN)
336 static inline bool
__xfs_rwsem_islocked(struct rw_semaphore * rwsem,bool shared)337 __xfs_rwsem_islocked(
338 struct rw_semaphore *rwsem,
339 bool shared)
340 {
341 if (!debug_locks)
342 return rwsem_is_locked(rwsem);
343
344 if (!shared)
345 return lockdep_is_held_type(rwsem, 0);
346
347 /*
348 * We are checking that the lock is held at least in shared
349 * mode but don't care that it might be held exclusively
350 * (i.e. shared | excl). Hence we check if the lock is held
351 * in any mode rather than an explicit shared mode.
352 */
353 return lockdep_is_held_type(rwsem, -1);
354 }
355
356 bool
xfs_isilocked(struct xfs_inode * ip,uint lock_flags)357 xfs_isilocked(
358 struct xfs_inode *ip,
359 uint lock_flags)
360 {
361 if (lock_flags & (XFS_ILOCK_EXCL|XFS_ILOCK_SHARED)) {
362 if (!(lock_flags & XFS_ILOCK_SHARED))
363 return !!ip->i_lock.mr_writer;
364 return rwsem_is_locked(&ip->i_lock.mr_lock);
365 }
366
367 if (lock_flags & (XFS_MMAPLOCK_EXCL|XFS_MMAPLOCK_SHARED)) {
368 return __xfs_rwsem_islocked(&VFS_I(ip)->i_mapping->invalidate_lock,
369 (lock_flags & XFS_MMAPLOCK_SHARED));
370 }
371
372 if (lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) {
373 return __xfs_rwsem_islocked(&VFS_I(ip)->i_rwsem,
374 (lock_flags & XFS_IOLOCK_SHARED));
375 }
376
377 ASSERT(0);
378 return false;
379 }
380 #endif
381
382 /*
383 * xfs_lockdep_subclass_ok() is only used in an ASSERT, so is only called when
384 * DEBUG or XFS_WARN is set. And MAX_LOCKDEP_SUBCLASSES is then only defined
385 * when CONFIG_LOCKDEP is set. Hence the complex define below to avoid build
386 * errors and warnings.
387 */
388 #if (defined(DEBUG) || defined(XFS_WARN)) && defined(CONFIG_LOCKDEP)
389 static bool
xfs_lockdep_subclass_ok(int subclass)390 xfs_lockdep_subclass_ok(
391 int subclass)
392 {
393 return subclass < MAX_LOCKDEP_SUBCLASSES;
394 }
395 #else
396 #define xfs_lockdep_subclass_ok(subclass) (true)
397 #endif
398
399 /*
400 * Bump the subclass so xfs_lock_inodes() acquires each lock with a different
401 * value. This can be called for any type of inode lock combination, including
402 * parent locking. Care must be taken to ensure we don't overrun the subclass
403 * storage fields in the class mask we build.
404 */
405 static inline uint
xfs_lock_inumorder(uint lock_mode,uint subclass)406 xfs_lock_inumorder(
407 uint lock_mode,
408 uint subclass)
409 {
410 uint class = 0;
411
412 ASSERT(!(lock_mode & (XFS_ILOCK_PARENT | XFS_ILOCK_RTBITMAP |
413 XFS_ILOCK_RTSUM)));
414 ASSERT(xfs_lockdep_subclass_ok(subclass));
415
416 if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)) {
417 ASSERT(subclass <= XFS_IOLOCK_MAX_SUBCLASS);
418 class += subclass << XFS_IOLOCK_SHIFT;
419 }
420
421 if (lock_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) {
422 ASSERT(subclass <= XFS_MMAPLOCK_MAX_SUBCLASS);
423 class += subclass << XFS_MMAPLOCK_SHIFT;
424 }
425
426 if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) {
427 ASSERT(subclass <= XFS_ILOCK_MAX_SUBCLASS);
428 class += subclass << XFS_ILOCK_SHIFT;
429 }
430
431 return (lock_mode & ~XFS_LOCK_SUBCLASS_MASK) | class;
432 }
433
434 /*
435 * The following routine will lock n inodes in exclusive mode. We assume the
436 * caller calls us with the inodes in i_ino order.
437 *
438 * We need to detect deadlock where an inode that we lock is in the AIL and we
439 * start waiting for another inode that is locked by a thread in a long running
440 * transaction (such as truncate). This can result in deadlock since the long
441 * running trans might need to wait for the inode we just locked in order to
442 * push the tail and free space in the log.
443 *
444 * xfs_lock_inodes() can only be used to lock one type of lock at a time -
445 * the iolock, the mmaplock or the ilock, but not more than one at a time. If we
446 * lock more than one at a time, lockdep will report false positives saying we
447 * have violated locking orders.
448 */
449 static void
xfs_lock_inodes(struct xfs_inode ** ips,int inodes,uint lock_mode)450 xfs_lock_inodes(
451 struct xfs_inode **ips,
452 int inodes,
453 uint lock_mode)
454 {
455 int attempts = 0;
456 uint i;
457 int j;
458 bool try_lock;
459 struct xfs_log_item *lp;
460
461 /*
462 * Currently supports between 2 and 5 inodes with exclusive locking. We
463 * support an arbitrary depth of locking here, but absolute limits on
464 * inodes depend on the type of locking and the limits placed by
465 * lockdep annotations in xfs_lock_inumorder. These are all checked by
466 * the asserts.
467 */
468 ASSERT(ips && inodes >= 2 && inodes <= 5);
469 ASSERT(lock_mode & (XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL |
470 XFS_ILOCK_EXCL));
471 ASSERT(!(lock_mode & (XFS_IOLOCK_SHARED | XFS_MMAPLOCK_SHARED |
472 XFS_ILOCK_SHARED)));
473 ASSERT(!(lock_mode & XFS_MMAPLOCK_EXCL) ||
474 inodes <= XFS_MMAPLOCK_MAX_SUBCLASS + 1);
475 ASSERT(!(lock_mode & XFS_ILOCK_EXCL) ||
476 inodes <= XFS_ILOCK_MAX_SUBCLASS + 1);
477
478 if (lock_mode & XFS_IOLOCK_EXCL) {
479 ASSERT(!(lock_mode & (XFS_MMAPLOCK_EXCL | XFS_ILOCK_EXCL)));
480 } else if (lock_mode & XFS_MMAPLOCK_EXCL)
481 ASSERT(!(lock_mode & XFS_ILOCK_EXCL));
482
483 again:
484 try_lock = false;
485 i = 0;
486 for (; i < inodes; i++) {
487 ASSERT(ips[i]);
488
489 if (i && (ips[i] == ips[i - 1])) /* Already locked */
490 continue;
491
492 /*
493 * If try_lock is not set yet, make sure all locked inodes are
494 * not in the AIL. If any are, set try_lock to be used later.
495 */
496 if (!try_lock) {
497 for (j = (i - 1); j >= 0 && !try_lock; j--) {
498 lp = &ips[j]->i_itemp->ili_item;
499 if (lp && test_bit(XFS_LI_IN_AIL, &lp->li_flags))
500 try_lock = true;
501 }
502 }
503
504 /*
505 * If any of the previous locks we have locked is in the AIL,
506 * we must TRY to get the second and subsequent locks. If
507 * we can't get any, we must release all we have
508 * and try again.
509 */
510 if (!try_lock) {
511 xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
512 continue;
513 }
514
515 /* try_lock means we have an inode locked that is in the AIL. */
516 ASSERT(i != 0);
517 if (xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i)))
518 continue;
519
520 /*
521 * Unlock all previous guys and try again. xfs_iunlock will try
522 * to push the tail if the inode is in the AIL.
523 */
524 attempts++;
525 for (j = i - 1; j >= 0; j--) {
526 /*
527 * Check to see if we've already unlocked this one. Not
528 * the first one going back, and the inode ptr is the
529 * same.
530 */
531 if (j != (i - 1) && ips[j] == ips[j + 1])
532 continue;
533
534 xfs_iunlock(ips[j], lock_mode);
535 }
536
537 if ((attempts % 5) == 0) {
538 delay(1); /* Don't just spin the CPU */
539 }
540 goto again;
541 }
542 }
543
544 /*
545 * xfs_lock_two_inodes() can only be used to lock ilock. The iolock and
546 * mmaplock must be double-locked separately since we use i_rwsem and
547 * invalidate_lock for that. We now support taking one lock EXCL and the
548 * other SHARED.
549 */
550 void
xfs_lock_two_inodes(struct xfs_inode * ip0,uint ip0_mode,struct xfs_inode * ip1,uint ip1_mode)551 xfs_lock_two_inodes(
552 struct xfs_inode *ip0,
553 uint ip0_mode,
554 struct xfs_inode *ip1,
555 uint ip1_mode)
556 {
557 int attempts = 0;
558 struct xfs_log_item *lp;
559
560 ASSERT(hweight32(ip0_mode) == 1);
561 ASSERT(hweight32(ip1_mode) == 1);
562 ASSERT(!(ip0_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
563 ASSERT(!(ip1_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
564 ASSERT(!(ip0_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)));
565 ASSERT(!(ip1_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)));
566 ASSERT(ip0->i_ino != ip1->i_ino);
567
568 if (ip0->i_ino > ip1->i_ino) {
569 swap(ip0, ip1);
570 swap(ip0_mode, ip1_mode);
571 }
572
573 again:
574 xfs_ilock(ip0, xfs_lock_inumorder(ip0_mode, 0));
575
576 /*
577 * If the first lock we have locked is in the AIL, we must TRY to get
578 * the second lock. If we can't get it, we must release the first one
579 * and try again.
580 */
581 lp = &ip0->i_itemp->ili_item;
582 if (lp && test_bit(XFS_LI_IN_AIL, &lp->li_flags)) {
583 if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(ip1_mode, 1))) {
584 xfs_iunlock(ip0, ip0_mode);
585 if ((++attempts % 5) == 0)
586 delay(1); /* Don't just spin the CPU */
587 goto again;
588 }
589 } else {
590 xfs_ilock(ip1, xfs_lock_inumorder(ip1_mode, 1));
591 }
592 }
593
594 uint
xfs_ip2xflags(struct xfs_inode * ip)595 xfs_ip2xflags(
596 struct xfs_inode *ip)
597 {
598 uint flags = 0;
599
600 if (ip->i_diflags & XFS_DIFLAG_ANY) {
601 if (ip->i_diflags & XFS_DIFLAG_REALTIME)
602 flags |= FS_XFLAG_REALTIME;
603 if (ip->i_diflags & XFS_DIFLAG_PREALLOC)
604 flags |= FS_XFLAG_PREALLOC;
605 if (ip->i_diflags & XFS_DIFLAG_IMMUTABLE)
606 flags |= FS_XFLAG_IMMUTABLE;
607 if (ip->i_diflags & XFS_DIFLAG_APPEND)
608 flags |= FS_XFLAG_APPEND;
609 if (ip->i_diflags & XFS_DIFLAG_SYNC)
610 flags |= FS_XFLAG_SYNC;
611 if (ip->i_diflags & XFS_DIFLAG_NOATIME)
612 flags |= FS_XFLAG_NOATIME;
613 if (ip->i_diflags & XFS_DIFLAG_NODUMP)
614 flags |= FS_XFLAG_NODUMP;
615 if (ip->i_diflags & XFS_DIFLAG_RTINHERIT)
616 flags |= FS_XFLAG_RTINHERIT;
617 if (ip->i_diflags & XFS_DIFLAG_PROJINHERIT)
618 flags |= FS_XFLAG_PROJINHERIT;
619 if (ip->i_diflags & XFS_DIFLAG_NOSYMLINKS)
620 flags |= FS_XFLAG_NOSYMLINKS;
621 if (ip->i_diflags & XFS_DIFLAG_EXTSIZE)
622 flags |= FS_XFLAG_EXTSIZE;
623 if (ip->i_diflags & XFS_DIFLAG_EXTSZINHERIT)
624 flags |= FS_XFLAG_EXTSZINHERIT;
625 if (ip->i_diflags & XFS_DIFLAG_NODEFRAG)
626 flags |= FS_XFLAG_NODEFRAG;
627 if (ip->i_diflags & XFS_DIFLAG_FILESTREAM)
628 flags |= FS_XFLAG_FILESTREAM;
629 }
630
631 if (ip->i_diflags2 & XFS_DIFLAG2_ANY) {
632 if (ip->i_diflags2 & XFS_DIFLAG2_DAX)
633 flags |= FS_XFLAG_DAX;
634 if (ip->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE)
635 flags |= FS_XFLAG_COWEXTSIZE;
636 }
637
638 if (XFS_IFORK_Q(ip))
639 flags |= FS_XFLAG_HASATTR;
640 return flags;
641 }
642
643 /*
644 * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
645 * is allowed, otherwise it has to be an exact match. If a CI match is found,
646 * ci_name->name will point to a the actual name (caller must free) or
647 * will be set to NULL if an exact match is found.
648 */
649 int
xfs_lookup(struct xfs_inode * dp,const struct xfs_name * name,struct xfs_inode ** ipp,struct xfs_name * ci_name)650 xfs_lookup(
651 struct xfs_inode *dp,
652 const struct xfs_name *name,
653 struct xfs_inode **ipp,
654 struct xfs_name *ci_name)
655 {
656 xfs_ino_t inum;
657 int error;
658
659 trace_xfs_lookup(dp, name);
660
661 if (xfs_is_shutdown(dp->i_mount))
662 return -EIO;
663
664 error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
665 if (error)
666 goto out_unlock;
667
668 error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp);
669 if (error)
670 goto out_free_name;
671
672 return 0;
673
674 out_free_name:
675 if (ci_name)
676 kmem_free(ci_name->name);
677 out_unlock:
678 *ipp = NULL;
679 return error;
680 }
681
682 /* Propagate di_flags from a parent inode to a child inode. */
683 static void
xfs_inode_inherit_flags(struct xfs_inode * ip,const struct xfs_inode * pip)684 xfs_inode_inherit_flags(
685 struct xfs_inode *ip,
686 const struct xfs_inode *pip)
687 {
688 unsigned int di_flags = 0;
689 xfs_failaddr_t failaddr;
690 umode_t mode = VFS_I(ip)->i_mode;
691
692 if (S_ISDIR(mode)) {
693 if (pip->i_diflags & XFS_DIFLAG_RTINHERIT)
694 di_flags |= XFS_DIFLAG_RTINHERIT;
695 if (pip->i_diflags & XFS_DIFLAG_EXTSZINHERIT) {
696 di_flags |= XFS_DIFLAG_EXTSZINHERIT;
697 ip->i_extsize = pip->i_extsize;
698 }
699 if (pip->i_diflags & XFS_DIFLAG_PROJINHERIT)
700 di_flags |= XFS_DIFLAG_PROJINHERIT;
701 } else if (S_ISREG(mode)) {
702 if ((pip->i_diflags & XFS_DIFLAG_RTINHERIT) &&
703 xfs_has_realtime(ip->i_mount))
704 di_flags |= XFS_DIFLAG_REALTIME;
705 if (pip->i_diflags & XFS_DIFLAG_EXTSZINHERIT) {
706 di_flags |= XFS_DIFLAG_EXTSIZE;
707 ip->i_extsize = pip->i_extsize;
708 }
709 }
710 if ((pip->i_diflags & XFS_DIFLAG_NOATIME) &&
711 xfs_inherit_noatime)
712 di_flags |= XFS_DIFLAG_NOATIME;
713 if ((pip->i_diflags & XFS_DIFLAG_NODUMP) &&
714 xfs_inherit_nodump)
715 di_flags |= XFS_DIFLAG_NODUMP;
716 if ((pip->i_diflags & XFS_DIFLAG_SYNC) &&
717 xfs_inherit_sync)
718 di_flags |= XFS_DIFLAG_SYNC;
719 if ((pip->i_diflags & XFS_DIFLAG_NOSYMLINKS) &&
720 xfs_inherit_nosymlinks)
721 di_flags |= XFS_DIFLAG_NOSYMLINKS;
722 if ((pip->i_diflags & XFS_DIFLAG_NODEFRAG) &&
723 xfs_inherit_nodefrag)
724 di_flags |= XFS_DIFLAG_NODEFRAG;
725 if (pip->i_diflags & XFS_DIFLAG_FILESTREAM)
726 di_flags |= XFS_DIFLAG_FILESTREAM;
727
728 ip->i_diflags |= di_flags;
729
730 /*
731 * Inode verifiers on older kernels only check that the extent size
732 * hint is an integer multiple of the rt extent size on realtime files.
733 * They did not check the hint alignment on a directory with both
734 * rtinherit and extszinherit flags set. If the misaligned hint is
735 * propagated from a directory into a new realtime file, new file
736 * allocations will fail due to math errors in the rt allocator and/or
737 * trip the verifiers. Validate the hint settings in the new file so
738 * that we don't let broken hints propagate.
739 */
740 failaddr = xfs_inode_validate_extsize(ip->i_mount, ip->i_extsize,
741 VFS_I(ip)->i_mode, ip->i_diflags);
742 if (failaddr) {
743 ip->i_diflags &= ~(XFS_DIFLAG_EXTSIZE |
744 XFS_DIFLAG_EXTSZINHERIT);
745 ip->i_extsize = 0;
746 }
747 }
748
749 /* Propagate di_flags2 from a parent inode to a child inode. */
750 static void
xfs_inode_inherit_flags2(struct xfs_inode * ip,const struct xfs_inode * pip)751 xfs_inode_inherit_flags2(
752 struct xfs_inode *ip,
753 const struct xfs_inode *pip)
754 {
755 xfs_failaddr_t failaddr;
756
757 if (pip->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE) {
758 ip->i_diflags2 |= XFS_DIFLAG2_COWEXTSIZE;
759 ip->i_cowextsize = pip->i_cowextsize;
760 }
761 if (pip->i_diflags2 & XFS_DIFLAG2_DAX)
762 ip->i_diflags2 |= XFS_DIFLAG2_DAX;
763
764 /* Don't let invalid cowextsize hints propagate. */
765 failaddr = xfs_inode_validate_cowextsize(ip->i_mount, ip->i_cowextsize,
766 VFS_I(ip)->i_mode, ip->i_diflags, ip->i_diflags2);
767 if (failaddr) {
768 ip->i_diflags2 &= ~XFS_DIFLAG2_COWEXTSIZE;
769 ip->i_cowextsize = 0;
770 }
771 }
772
773 /*
774 * Initialise a newly allocated inode and return the in-core inode to the
775 * caller locked exclusively.
776 */
777 int
xfs_init_new_inode(struct user_namespace * mnt_userns,struct xfs_trans * tp,struct xfs_inode * pip,xfs_ino_t ino,umode_t mode,xfs_nlink_t nlink,dev_t rdev,prid_t prid,bool init_xattrs,struct xfs_inode ** ipp)778 xfs_init_new_inode(
779 struct user_namespace *mnt_userns,
780 struct xfs_trans *tp,
781 struct xfs_inode *pip,
782 xfs_ino_t ino,
783 umode_t mode,
784 xfs_nlink_t nlink,
785 dev_t rdev,
786 prid_t prid,
787 bool init_xattrs,
788 struct xfs_inode **ipp)
789 {
790 struct inode *dir = pip ? VFS_I(pip) : NULL;
791 struct xfs_mount *mp = tp->t_mountp;
792 struct xfs_inode *ip;
793 unsigned int flags;
794 int error;
795 struct timespec64 tv;
796 struct inode *inode;
797
798 /*
799 * Protect against obviously corrupt allocation btree records. Later
800 * xfs_iget checks will catch re-allocation of other active in-memory
801 * and on-disk inodes. If we don't catch reallocating the parent inode
802 * here we will deadlock in xfs_iget() so we have to do these checks
803 * first.
804 */
805 if ((pip && ino == pip->i_ino) || !xfs_verify_dir_ino(mp, ino)) {
806 xfs_alert(mp, "Allocated a known in-use inode 0x%llx!", ino);
807 return -EFSCORRUPTED;
808 }
809
810 /*
811 * Get the in-core inode with the lock held exclusively to prevent
812 * others from looking at until we're done.
813 */
814 error = xfs_iget(mp, tp, ino, XFS_IGET_CREATE, XFS_ILOCK_EXCL, &ip);
815 if (error)
816 return error;
817
818 ASSERT(ip != NULL);
819 inode = VFS_I(ip);
820 set_nlink(inode, nlink);
821 inode->i_rdev = rdev;
822 ip->i_projid = prid;
823
824 if (dir && !(dir->i_mode & S_ISGID) && xfs_has_grpid(mp)) {
825 inode_fsuid_set(inode, mnt_userns);
826 inode->i_gid = dir->i_gid;
827 inode->i_mode = mode;
828 } else {
829 inode_init_owner(mnt_userns, inode, dir, mode);
830 }
831
832 /*
833 * If the group ID of the new file does not match the effective group
834 * ID or one of the supplementary group IDs, the S_ISGID bit is cleared
835 * (and only if the irix_sgid_inherit compatibility variable is set).
836 */
837 if (irix_sgid_inherit &&
838 (inode->i_mode & S_ISGID) &&
839 !in_group_p(i_gid_into_mnt(mnt_userns, inode)))
840 inode->i_mode &= ~S_ISGID;
841
842 ip->i_disk_size = 0;
843 ip->i_df.if_nextents = 0;
844 ASSERT(ip->i_nblocks == 0);
845
846 tv = current_time(inode);
847 inode->i_mtime = tv;
848 inode->i_atime = tv;
849 inode->i_ctime = tv;
850
851 ip->i_extsize = 0;
852 ip->i_diflags = 0;
853
854 if (xfs_has_v3inodes(mp)) {
855 inode_set_iversion(inode, 1);
856 ip->i_cowextsize = 0;
857 ip->i_crtime = tv;
858 }
859
860 flags = XFS_ILOG_CORE;
861 switch (mode & S_IFMT) {
862 case S_IFIFO:
863 case S_IFCHR:
864 case S_IFBLK:
865 case S_IFSOCK:
866 ip->i_df.if_format = XFS_DINODE_FMT_DEV;
867 flags |= XFS_ILOG_DEV;
868 break;
869 case S_IFREG:
870 case S_IFDIR:
871 if (pip && (pip->i_diflags & XFS_DIFLAG_ANY))
872 xfs_inode_inherit_flags(ip, pip);
873 if (pip && (pip->i_diflags2 & XFS_DIFLAG2_ANY))
874 xfs_inode_inherit_flags2(ip, pip);
875 fallthrough;
876 case S_IFLNK:
877 ip->i_df.if_format = XFS_DINODE_FMT_EXTENTS;
878 ip->i_df.if_bytes = 0;
879 ip->i_df.if_u1.if_root = NULL;
880 break;
881 default:
882 ASSERT(0);
883 }
884
885 /*
886 * If we need to create attributes immediately after allocating the
887 * inode, initialise an empty attribute fork right now. We use the
888 * default fork offset for attributes here as we don't know exactly what
889 * size or how many attributes we might be adding. We can do this
890 * safely here because we know the data fork is completely empty and
891 * this saves us from needing to run a separate transaction to set the
892 * fork offset in the immediate future.
893 */
894 if (init_xattrs && xfs_has_attr(mp)) {
895 ip->i_forkoff = xfs_default_attroffset(ip) >> 3;
896 ip->i_afp = xfs_ifork_alloc(XFS_DINODE_FMT_EXTENTS, 0);
897 }
898
899 /*
900 * Log the new values stuffed into the inode.
901 */
902 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
903 xfs_trans_log_inode(tp, ip, flags);
904
905 /* now that we have an i_mode we can setup the inode structure */
906 xfs_setup_inode(ip);
907
908 *ipp = ip;
909 return 0;
910 }
911
912 /*
913 * Decrement the link count on an inode & log the change. If this causes the
914 * link count to go to zero, move the inode to AGI unlinked list so that it can
915 * be freed when the last active reference goes away via xfs_inactive().
916 */
917 static int /* error */
xfs_droplink(xfs_trans_t * tp,xfs_inode_t * ip)918 xfs_droplink(
919 xfs_trans_t *tp,
920 xfs_inode_t *ip)
921 {
922 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
923
924 drop_nlink(VFS_I(ip));
925 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
926
927 if (VFS_I(ip)->i_nlink)
928 return 0;
929
930 return xfs_iunlink(tp, ip);
931 }
932
933 /*
934 * Increment the link count on an inode & log the change.
935 */
936 static void
xfs_bumplink(xfs_trans_t * tp,xfs_inode_t * ip)937 xfs_bumplink(
938 xfs_trans_t *tp,
939 xfs_inode_t *ip)
940 {
941 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_CHG);
942
943 inc_nlink(VFS_I(ip));
944 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
945 }
946
947 int
xfs_create(struct user_namespace * mnt_userns,xfs_inode_t * dp,struct xfs_name * name,umode_t mode,dev_t rdev,bool init_xattrs,xfs_inode_t ** ipp)948 xfs_create(
949 struct user_namespace *mnt_userns,
950 xfs_inode_t *dp,
951 struct xfs_name *name,
952 umode_t mode,
953 dev_t rdev,
954 bool init_xattrs,
955 xfs_inode_t **ipp)
956 {
957 int is_dir = S_ISDIR(mode);
958 struct xfs_mount *mp = dp->i_mount;
959 struct xfs_inode *ip = NULL;
960 struct xfs_trans *tp = NULL;
961 int error;
962 bool unlock_dp_on_error = false;
963 prid_t prid;
964 struct xfs_dquot *udqp = NULL;
965 struct xfs_dquot *gdqp = NULL;
966 struct xfs_dquot *pdqp = NULL;
967 struct xfs_trans_res *tres;
968 uint resblks;
969 xfs_ino_t ino;
970
971 trace_xfs_create(dp, name);
972
973 if (xfs_is_shutdown(mp))
974 return -EIO;
975
976 prid = xfs_get_initial_prid(dp);
977
978 /*
979 * Make sure that we have allocated dquot(s) on disk.
980 */
981 error = xfs_qm_vop_dqalloc(dp, mapped_fsuid(mnt_userns, &init_user_ns),
982 mapped_fsgid(mnt_userns, &init_user_ns), prid,
983 XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
984 &udqp, &gdqp, &pdqp);
985 if (error)
986 return error;
987
988 if (is_dir) {
989 resblks = XFS_MKDIR_SPACE_RES(mp, name->len);
990 tres = &M_RES(mp)->tr_mkdir;
991 } else {
992 resblks = XFS_CREATE_SPACE_RES(mp, name->len);
993 tres = &M_RES(mp)->tr_create;
994 }
995
996 /*
997 * Initially assume that the file does not exist and
998 * reserve the resources for that case. If that is not
999 * the case we'll drop the one we have and get a more
1000 * appropriate transaction later.
1001 */
1002 error = xfs_trans_alloc_icreate(mp, tres, udqp, gdqp, pdqp, resblks,
1003 &tp);
1004 if (error == -ENOSPC) {
1005 /* flush outstanding delalloc blocks and retry */
1006 xfs_flush_inodes(mp);
1007 error = xfs_trans_alloc_icreate(mp, tres, udqp, gdqp, pdqp,
1008 resblks, &tp);
1009 }
1010 if (error)
1011 goto out_release_dquots;
1012
1013 xfs_ilock(dp, XFS_ILOCK_EXCL | XFS_ILOCK_PARENT);
1014 unlock_dp_on_error = true;
1015
1016 /*
1017 * A newly created regular or special file just has one directory
1018 * entry pointing to them, but a directory also the "." entry
1019 * pointing to itself.
1020 */
1021 error = xfs_dialloc(&tp, dp->i_ino, mode, &ino);
1022 if (!error)
1023 error = xfs_init_new_inode(mnt_userns, tp, dp, ino, mode,
1024 is_dir ? 2 : 1, rdev, prid, init_xattrs, &ip);
1025 if (error)
1026 goto out_trans_cancel;
1027
1028 /*
1029 * Now we join the directory inode to the transaction. We do not do it
1030 * earlier because xfs_dialloc might commit the previous transaction
1031 * (and release all the locks). An error from here on will result in
1032 * the transaction cancel unlocking dp so don't do it explicitly in the
1033 * error path.
1034 */
1035 xfs_trans_ijoin(tp, dp, XFS_ILOCK_EXCL);
1036 unlock_dp_on_error = false;
1037
1038 error = xfs_dir_createname(tp, dp, name, ip->i_ino,
1039 resblks - XFS_IALLOC_SPACE_RES(mp));
1040 if (error) {
1041 ASSERT(error != -ENOSPC);
1042 goto out_trans_cancel;
1043 }
1044 xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1045 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
1046
1047 if (is_dir) {
1048 error = xfs_dir_init(tp, ip, dp);
1049 if (error)
1050 goto out_trans_cancel;
1051
1052 xfs_bumplink(tp, dp);
1053 }
1054
1055 /*
1056 * If this is a synchronous mount, make sure that the
1057 * create transaction goes to disk before returning to
1058 * the user.
1059 */
1060 if (xfs_has_wsync(mp) || xfs_has_dirsync(mp))
1061 xfs_trans_set_sync(tp);
1062
1063 /*
1064 * Attach the dquot(s) to the inodes and modify them incore.
1065 * These ids of the inode couldn't have changed since the new
1066 * inode has been locked ever since it was created.
1067 */
1068 xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1069
1070 error = xfs_trans_commit(tp);
1071 if (error)
1072 goto out_release_inode;
1073
1074 xfs_qm_dqrele(udqp);
1075 xfs_qm_dqrele(gdqp);
1076 xfs_qm_dqrele(pdqp);
1077
1078 *ipp = ip;
1079 return 0;
1080
1081 out_trans_cancel:
1082 xfs_trans_cancel(tp);
1083 out_release_inode:
1084 /*
1085 * Wait until after the current transaction is aborted to finish the
1086 * setup of the inode and release the inode. This prevents recursive
1087 * transactions and deadlocks from xfs_inactive.
1088 */
1089 if (ip) {
1090 xfs_finish_inode_setup(ip);
1091 xfs_irele(ip);
1092 }
1093 out_release_dquots:
1094 xfs_qm_dqrele(udqp);
1095 xfs_qm_dqrele(gdqp);
1096 xfs_qm_dqrele(pdqp);
1097
1098 if (unlock_dp_on_error)
1099 xfs_iunlock(dp, XFS_ILOCK_EXCL);
1100 return error;
1101 }
1102
1103 int
xfs_create_tmpfile(struct user_namespace * mnt_userns,struct xfs_inode * dp,umode_t mode,struct xfs_inode ** ipp)1104 xfs_create_tmpfile(
1105 struct user_namespace *mnt_userns,
1106 struct xfs_inode *dp,
1107 umode_t mode,
1108 struct xfs_inode **ipp)
1109 {
1110 struct xfs_mount *mp = dp->i_mount;
1111 struct xfs_inode *ip = NULL;
1112 struct xfs_trans *tp = NULL;
1113 int error;
1114 prid_t prid;
1115 struct xfs_dquot *udqp = NULL;
1116 struct xfs_dquot *gdqp = NULL;
1117 struct xfs_dquot *pdqp = NULL;
1118 struct xfs_trans_res *tres;
1119 uint resblks;
1120 xfs_ino_t ino;
1121
1122 if (xfs_is_shutdown(mp))
1123 return -EIO;
1124
1125 prid = xfs_get_initial_prid(dp);
1126
1127 /*
1128 * Make sure that we have allocated dquot(s) on disk.
1129 */
1130 error = xfs_qm_vop_dqalloc(dp, mapped_fsuid(mnt_userns, &init_user_ns),
1131 mapped_fsgid(mnt_userns, &init_user_ns), prid,
1132 XFS_QMOPT_QUOTALL | XFS_QMOPT_INHERIT,
1133 &udqp, &gdqp, &pdqp);
1134 if (error)
1135 return error;
1136
1137 resblks = XFS_IALLOC_SPACE_RES(mp);
1138 tres = &M_RES(mp)->tr_create_tmpfile;
1139
1140 error = xfs_trans_alloc_icreate(mp, tres, udqp, gdqp, pdqp, resblks,
1141 &tp);
1142 if (error)
1143 goto out_release_dquots;
1144
1145 error = xfs_dialloc(&tp, dp->i_ino, mode, &ino);
1146 if (!error)
1147 error = xfs_init_new_inode(mnt_userns, tp, dp, ino, mode,
1148 0, 0, prid, false, &ip);
1149 if (error)
1150 goto out_trans_cancel;
1151
1152 if (xfs_has_wsync(mp))
1153 xfs_trans_set_sync(tp);
1154
1155 /*
1156 * Attach the dquot(s) to the inodes and modify them incore.
1157 * These ids of the inode couldn't have changed since the new
1158 * inode has been locked ever since it was created.
1159 */
1160 xfs_qm_vop_create_dqattach(tp, ip, udqp, gdqp, pdqp);
1161
1162 error = xfs_iunlink(tp, ip);
1163 if (error)
1164 goto out_trans_cancel;
1165
1166 error = xfs_trans_commit(tp);
1167 if (error)
1168 goto out_release_inode;
1169
1170 xfs_qm_dqrele(udqp);
1171 xfs_qm_dqrele(gdqp);
1172 xfs_qm_dqrele(pdqp);
1173
1174 *ipp = ip;
1175 return 0;
1176
1177 out_trans_cancel:
1178 xfs_trans_cancel(tp);
1179 out_release_inode:
1180 /*
1181 * Wait until after the current transaction is aborted to finish the
1182 * setup of the inode and release the inode. This prevents recursive
1183 * transactions and deadlocks from xfs_inactive.
1184 */
1185 if (ip) {
1186 xfs_finish_inode_setup(ip);
1187 xfs_irele(ip);
1188 }
1189 out_release_dquots:
1190 xfs_qm_dqrele(udqp);
1191 xfs_qm_dqrele(gdqp);
1192 xfs_qm_dqrele(pdqp);
1193
1194 return error;
1195 }
1196
1197 int
xfs_link(xfs_inode_t * tdp,xfs_inode_t * sip,struct xfs_name * target_name)1198 xfs_link(
1199 xfs_inode_t *tdp,
1200 xfs_inode_t *sip,
1201 struct xfs_name *target_name)
1202 {
1203 xfs_mount_t *mp = tdp->i_mount;
1204 xfs_trans_t *tp;
1205 int error, nospace_error = 0;
1206 int resblks;
1207
1208 trace_xfs_link(tdp, target_name);
1209
1210 ASSERT(!S_ISDIR(VFS_I(sip)->i_mode));
1211
1212 if (xfs_is_shutdown(mp))
1213 return -EIO;
1214
1215 error = xfs_qm_dqattach(sip);
1216 if (error)
1217 goto std_return;
1218
1219 error = xfs_qm_dqattach(tdp);
1220 if (error)
1221 goto std_return;
1222
1223 resblks = XFS_LINK_SPACE_RES(mp, target_name->len);
1224 error = xfs_trans_alloc_dir(tdp, &M_RES(mp)->tr_link, sip, &resblks,
1225 &tp, &nospace_error);
1226 if (error)
1227 goto std_return;
1228
1229 /*
1230 * If we are using project inheritance, we only allow hard link
1231 * creation in our tree when the project IDs are the same; else
1232 * the tree quota mechanism could be circumvented.
1233 */
1234 if (unlikely((tdp->i_diflags & XFS_DIFLAG_PROJINHERIT) &&
1235 tdp->i_projid != sip->i_projid)) {
1236 error = -EXDEV;
1237 goto error_return;
1238 }
1239
1240 if (!resblks) {
1241 error = xfs_dir_canenter(tp, tdp, target_name);
1242 if (error)
1243 goto error_return;
1244 }
1245
1246 /*
1247 * Handle initial link state of O_TMPFILE inode
1248 */
1249 if (VFS_I(sip)->i_nlink == 0) {
1250 struct xfs_perag *pag;
1251
1252 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, sip->i_ino));
1253 error = xfs_iunlink_remove(tp, pag, sip);
1254 xfs_perag_put(pag);
1255 if (error)
1256 goto error_return;
1257 }
1258
1259 error = xfs_dir_createname(tp, tdp, target_name, sip->i_ino,
1260 resblks);
1261 if (error)
1262 goto error_return;
1263 xfs_trans_ichgtime(tp, tdp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1264 xfs_trans_log_inode(tp, tdp, XFS_ILOG_CORE);
1265
1266 xfs_bumplink(tp, sip);
1267
1268 /*
1269 * If this is a synchronous mount, make sure that the
1270 * link transaction goes to disk before returning to
1271 * the user.
1272 */
1273 if (xfs_has_wsync(mp) || xfs_has_dirsync(mp))
1274 xfs_trans_set_sync(tp);
1275
1276 return xfs_trans_commit(tp);
1277
1278 error_return:
1279 xfs_trans_cancel(tp);
1280 std_return:
1281 if (error == -ENOSPC && nospace_error)
1282 error = nospace_error;
1283 return error;
1284 }
1285
1286 /* Clear the reflink flag and the cowblocks tag if possible. */
1287 static void
xfs_itruncate_clear_reflink_flags(struct xfs_inode * ip)1288 xfs_itruncate_clear_reflink_flags(
1289 struct xfs_inode *ip)
1290 {
1291 struct xfs_ifork *dfork;
1292 struct xfs_ifork *cfork;
1293
1294 if (!xfs_is_reflink_inode(ip))
1295 return;
1296 dfork = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1297 cfork = XFS_IFORK_PTR(ip, XFS_COW_FORK);
1298 if (dfork->if_bytes == 0 && cfork->if_bytes == 0)
1299 ip->i_diflags2 &= ~XFS_DIFLAG2_REFLINK;
1300 if (cfork->if_bytes == 0)
1301 xfs_inode_clear_cowblocks_tag(ip);
1302 }
1303
1304 /*
1305 * Free up the underlying blocks past new_size. The new size must be smaller
1306 * than the current size. This routine can be used both for the attribute and
1307 * data fork, and does not modify the inode size, which is left to the caller.
1308 *
1309 * The transaction passed to this routine must have made a permanent log
1310 * reservation of at least XFS_ITRUNCATE_LOG_RES. This routine may commit the
1311 * given transaction and start new ones, so make sure everything involved in
1312 * the transaction is tidy before calling here. Some transaction will be
1313 * returned to the caller to be committed. The incoming transaction must
1314 * already include the inode, and both inode locks must be held exclusively.
1315 * The inode must also be "held" within the transaction. On return the inode
1316 * will be "held" within the returned transaction. This routine does NOT
1317 * require any disk space to be reserved for it within the transaction.
1318 *
1319 * If we get an error, we must return with the inode locked and linked into the
1320 * current transaction. This keeps things simple for the higher level code,
1321 * because it always knows that the inode is locked and held in the transaction
1322 * that returns to it whether errors occur or not. We don't mark the inode
1323 * dirty on error so that transactions can be easily aborted if possible.
1324 */
1325 int
xfs_itruncate_extents_flags(struct xfs_trans ** tpp,struct xfs_inode * ip,int whichfork,xfs_fsize_t new_size,int flags)1326 xfs_itruncate_extents_flags(
1327 struct xfs_trans **tpp,
1328 struct xfs_inode *ip,
1329 int whichfork,
1330 xfs_fsize_t new_size,
1331 int flags)
1332 {
1333 struct xfs_mount *mp = ip->i_mount;
1334 struct xfs_trans *tp = *tpp;
1335 xfs_fileoff_t first_unmap_block;
1336 xfs_filblks_t unmap_len;
1337 int error = 0;
1338
1339 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1340 ASSERT(!atomic_read(&VFS_I(ip)->i_count) ||
1341 xfs_isilocked(ip, XFS_IOLOCK_EXCL));
1342 ASSERT(new_size <= XFS_ISIZE(ip));
1343 ASSERT(tp->t_flags & XFS_TRANS_PERM_LOG_RES);
1344 ASSERT(ip->i_itemp != NULL);
1345 ASSERT(ip->i_itemp->ili_lock_flags == 0);
1346 ASSERT(!XFS_NOT_DQATTACHED(mp, ip));
1347
1348 trace_xfs_itruncate_extents_start(ip, new_size);
1349
1350 flags |= xfs_bmapi_aflag(whichfork);
1351
1352 /*
1353 * Since it is possible for space to become allocated beyond
1354 * the end of the file (in a crash where the space is allocated
1355 * but the inode size is not yet updated), simply remove any
1356 * blocks which show up between the new EOF and the maximum
1357 * possible file size.
1358 *
1359 * We have to free all the blocks to the bmbt maximum offset, even if
1360 * the page cache can't scale that far.
1361 */
1362 first_unmap_block = XFS_B_TO_FSB(mp, (xfs_ufsize_t)new_size);
1363 if (!xfs_verify_fileoff(mp, first_unmap_block)) {
1364 WARN_ON_ONCE(first_unmap_block > XFS_MAX_FILEOFF);
1365 return 0;
1366 }
1367
1368 unmap_len = XFS_MAX_FILEOFF - first_unmap_block + 1;
1369 while (unmap_len > 0) {
1370 ASSERT(tp->t_firstblock == NULLFSBLOCK);
1371 error = __xfs_bunmapi(tp, ip, first_unmap_block, &unmap_len,
1372 flags, XFS_ITRUNC_MAX_EXTENTS);
1373 if (error)
1374 goto out;
1375
1376 /* free the just unmapped extents */
1377 error = xfs_defer_finish(&tp);
1378 if (error)
1379 goto out;
1380 }
1381
1382 if (whichfork == XFS_DATA_FORK) {
1383 /* Remove all pending CoW reservations. */
1384 error = xfs_reflink_cancel_cow_blocks(ip, &tp,
1385 first_unmap_block, XFS_MAX_FILEOFF, true);
1386 if (error)
1387 goto out;
1388
1389 xfs_itruncate_clear_reflink_flags(ip);
1390 }
1391
1392 /*
1393 * Always re-log the inode so that our permanent transaction can keep
1394 * on rolling it forward in the log.
1395 */
1396 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1397
1398 trace_xfs_itruncate_extents_end(ip, new_size);
1399
1400 out:
1401 *tpp = tp;
1402 return error;
1403 }
1404
1405 int
xfs_release(xfs_inode_t * ip)1406 xfs_release(
1407 xfs_inode_t *ip)
1408 {
1409 xfs_mount_t *mp = ip->i_mount;
1410 int error = 0;
1411
1412 if (!S_ISREG(VFS_I(ip)->i_mode) || (VFS_I(ip)->i_mode == 0))
1413 return 0;
1414
1415 /* If this is a read-only mount, don't do this (would generate I/O) */
1416 if (xfs_is_readonly(mp))
1417 return 0;
1418
1419 if (!xfs_is_shutdown(mp)) {
1420 int truncated;
1421
1422 /*
1423 * If we previously truncated this file and removed old data
1424 * in the process, we want to initiate "early" writeout on
1425 * the last close. This is an attempt to combat the notorious
1426 * NULL files problem which is particularly noticeable from a
1427 * truncate down, buffered (re-)write (delalloc), followed by
1428 * a crash. What we are effectively doing here is
1429 * significantly reducing the time window where we'd otherwise
1430 * be exposed to that problem.
1431 */
1432 truncated = xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED);
1433 if (truncated) {
1434 xfs_iflags_clear(ip, XFS_IDIRTY_RELEASE);
1435 if (ip->i_delayed_blks > 0) {
1436 error = filemap_flush(VFS_I(ip)->i_mapping);
1437 if (error)
1438 return error;
1439 }
1440 }
1441 }
1442
1443 if (VFS_I(ip)->i_nlink == 0)
1444 return 0;
1445
1446 /*
1447 * If we can't get the iolock just skip truncating the blocks past EOF
1448 * because we could deadlock with the mmap_lock otherwise. We'll get
1449 * another chance to drop them once the last reference to the inode is
1450 * dropped, so we'll never leak blocks permanently.
1451 */
1452 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL))
1453 return 0;
1454
1455 if (xfs_can_free_eofblocks(ip, false)) {
1456 /*
1457 * Check if the inode is being opened, written and closed
1458 * frequently and we have delayed allocation blocks outstanding
1459 * (e.g. streaming writes from the NFS server), truncating the
1460 * blocks past EOF will cause fragmentation to occur.
1461 *
1462 * In this case don't do the truncation, but we have to be
1463 * careful how we detect this case. Blocks beyond EOF show up as
1464 * i_delayed_blks even when the inode is clean, so we need to
1465 * truncate them away first before checking for a dirty release.
1466 * Hence on the first dirty close we will still remove the
1467 * speculative allocation, but after that we will leave it in
1468 * place.
1469 */
1470 if (xfs_iflags_test(ip, XFS_IDIRTY_RELEASE))
1471 goto out_unlock;
1472
1473 error = xfs_free_eofblocks(ip);
1474 if (error)
1475 goto out_unlock;
1476
1477 /* delalloc blocks after truncation means it really is dirty */
1478 if (ip->i_delayed_blks)
1479 xfs_iflags_set(ip, XFS_IDIRTY_RELEASE);
1480 }
1481
1482 out_unlock:
1483 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1484 return error;
1485 }
1486
1487 /*
1488 * xfs_inactive_truncate
1489 *
1490 * Called to perform a truncate when an inode becomes unlinked.
1491 */
1492 STATIC int
xfs_inactive_truncate(struct xfs_inode * ip)1493 xfs_inactive_truncate(
1494 struct xfs_inode *ip)
1495 {
1496 struct xfs_mount *mp = ip->i_mount;
1497 struct xfs_trans *tp;
1498 int error;
1499
1500 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_itruncate, 0, 0, 0, &tp);
1501 if (error) {
1502 ASSERT(xfs_is_shutdown(mp));
1503 return error;
1504 }
1505 xfs_ilock(ip, XFS_ILOCK_EXCL);
1506 xfs_trans_ijoin(tp, ip, 0);
1507
1508 /*
1509 * Log the inode size first to prevent stale data exposure in the event
1510 * of a system crash before the truncate completes. See the related
1511 * comment in xfs_vn_setattr_size() for details.
1512 */
1513 ip->i_disk_size = 0;
1514 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1515
1516 error = xfs_itruncate_extents(&tp, ip, XFS_DATA_FORK, 0);
1517 if (error)
1518 goto error_trans_cancel;
1519
1520 ASSERT(ip->i_df.if_nextents == 0);
1521
1522 error = xfs_trans_commit(tp);
1523 if (error)
1524 goto error_unlock;
1525
1526 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1527 return 0;
1528
1529 error_trans_cancel:
1530 xfs_trans_cancel(tp);
1531 error_unlock:
1532 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1533 return error;
1534 }
1535
1536 /*
1537 * xfs_inactive_ifree()
1538 *
1539 * Perform the inode free when an inode is unlinked.
1540 */
1541 STATIC int
xfs_inactive_ifree(struct xfs_inode * ip)1542 xfs_inactive_ifree(
1543 struct xfs_inode *ip)
1544 {
1545 struct xfs_mount *mp = ip->i_mount;
1546 struct xfs_trans *tp;
1547 int error;
1548
1549 /*
1550 * We try to use a per-AG reservation for any block needed by the finobt
1551 * tree, but as the finobt feature predates the per-AG reservation
1552 * support a degraded file system might not have enough space for the
1553 * reservation at mount time. In that case try to dip into the reserved
1554 * pool and pray.
1555 *
1556 * Send a warning if the reservation does happen to fail, as the inode
1557 * now remains allocated and sits on the unlinked list until the fs is
1558 * repaired.
1559 */
1560 if (unlikely(mp->m_finobt_nores)) {
1561 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree,
1562 XFS_IFREE_SPACE_RES(mp), 0, XFS_TRANS_RESERVE,
1563 &tp);
1564 } else {
1565 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ifree, 0, 0, 0, &tp);
1566 }
1567 if (error) {
1568 if (error == -ENOSPC) {
1569 xfs_warn_ratelimited(mp,
1570 "Failed to remove inode(s) from unlinked list. "
1571 "Please free space, unmount and run xfs_repair.");
1572 } else {
1573 ASSERT(xfs_is_shutdown(mp));
1574 }
1575 return error;
1576 }
1577
1578 /*
1579 * We do not hold the inode locked across the entire rolling transaction
1580 * here. We only need to hold it for the first transaction that
1581 * xfs_ifree() builds, which may mark the inode XFS_ISTALE if the
1582 * underlying cluster buffer is freed. Relogging an XFS_ISTALE inode
1583 * here breaks the relationship between cluster buffer invalidation and
1584 * stale inode invalidation on cluster buffer item journal commit
1585 * completion, and can result in leaving dirty stale inodes hanging
1586 * around in memory.
1587 *
1588 * We have no need for serialising this inode operation against other
1589 * operations - we freed the inode and hence reallocation is required
1590 * and that will serialise on reallocating the space the deferops need
1591 * to free. Hence we can unlock the inode on the first commit of
1592 * the transaction rather than roll it right through the deferops. This
1593 * avoids relogging the XFS_ISTALE inode.
1594 *
1595 * We check that xfs_ifree() hasn't grown an internal transaction roll
1596 * by asserting that the inode is still locked when it returns.
1597 */
1598 xfs_ilock(ip, XFS_ILOCK_EXCL);
1599 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
1600
1601 error = xfs_ifree(tp, ip);
1602 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
1603 if (error) {
1604 /*
1605 * If we fail to free the inode, shut down. The cancel
1606 * might do that, we need to make sure. Otherwise the
1607 * inode might be lost for a long time or forever.
1608 */
1609 if (!xfs_is_shutdown(mp)) {
1610 xfs_notice(mp, "%s: xfs_ifree returned error %d",
1611 __func__, error);
1612 xfs_force_shutdown(mp, SHUTDOWN_META_IO_ERROR);
1613 }
1614 xfs_trans_cancel(tp);
1615 return error;
1616 }
1617
1618 /*
1619 * Credit the quota account(s). The inode is gone.
1620 */
1621 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_ICOUNT, -1);
1622
1623 /*
1624 * Just ignore errors at this point. There is nothing we can do except
1625 * to try to keep going. Make sure it's not a silent error.
1626 */
1627 error = xfs_trans_commit(tp);
1628 if (error)
1629 xfs_notice(mp, "%s: xfs_trans_commit returned error %d",
1630 __func__, error);
1631
1632 return 0;
1633 }
1634
1635 /*
1636 * Returns true if we need to update the on-disk metadata before we can free
1637 * the memory used by this inode. Updates include freeing post-eof
1638 * preallocations; freeing COW staging extents; and marking the inode free in
1639 * the inobt if it is on the unlinked list.
1640 */
1641 bool
xfs_inode_needs_inactive(struct xfs_inode * ip)1642 xfs_inode_needs_inactive(
1643 struct xfs_inode *ip)
1644 {
1645 struct xfs_mount *mp = ip->i_mount;
1646 struct xfs_ifork *cow_ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
1647
1648 /*
1649 * If the inode is already free, then there can be nothing
1650 * to clean up here.
1651 */
1652 if (VFS_I(ip)->i_mode == 0)
1653 return false;
1654
1655 /* If this is a read-only mount, don't do this (would generate I/O) */
1656 if (xfs_is_readonly(mp))
1657 return false;
1658
1659 /* If the log isn't running, push inodes straight to reclaim. */
1660 if (xfs_is_shutdown(mp) || xfs_has_norecovery(mp))
1661 return false;
1662
1663 /* Metadata inodes require explicit resource cleanup. */
1664 if (xfs_is_metadata_inode(ip))
1665 return false;
1666
1667 /* Want to clean out the cow blocks if there are any. */
1668 if (cow_ifp && cow_ifp->if_bytes > 0)
1669 return true;
1670
1671 /* Unlinked files must be freed. */
1672 if (VFS_I(ip)->i_nlink == 0)
1673 return true;
1674
1675 /*
1676 * This file isn't being freed, so check if there are post-eof blocks
1677 * to free. @force is true because we are evicting an inode from the
1678 * cache. Post-eof blocks must be freed, lest we end up with broken
1679 * free space accounting.
1680 *
1681 * Note: don't bother with iolock here since lockdep complains about
1682 * acquiring it in reclaim context. We have the only reference to the
1683 * inode at this point anyways.
1684 */
1685 return xfs_can_free_eofblocks(ip, true);
1686 }
1687
1688 /*
1689 * xfs_inactive
1690 *
1691 * This is called when the vnode reference count for the vnode
1692 * goes to zero. If the file has been unlinked, then it must
1693 * now be truncated. Also, we clear all of the read-ahead state
1694 * kept for the inode here since the file is now closed.
1695 */
1696 void
xfs_inactive(xfs_inode_t * ip)1697 xfs_inactive(
1698 xfs_inode_t *ip)
1699 {
1700 struct xfs_mount *mp;
1701 int error;
1702 int truncate = 0;
1703
1704 /*
1705 * If the inode is already free, then there can be nothing
1706 * to clean up here.
1707 */
1708 if (VFS_I(ip)->i_mode == 0) {
1709 ASSERT(ip->i_df.if_broot_bytes == 0);
1710 goto out;
1711 }
1712
1713 mp = ip->i_mount;
1714 ASSERT(!xfs_iflags_test(ip, XFS_IRECOVERY));
1715
1716 /* If this is a read-only mount, don't do this (would generate I/O) */
1717 if (xfs_is_readonly(mp))
1718 goto out;
1719
1720 /* Metadata inodes require explicit resource cleanup. */
1721 if (xfs_is_metadata_inode(ip))
1722 goto out;
1723
1724 /* Try to clean out the cow blocks if there are any. */
1725 if (xfs_inode_has_cow_data(ip))
1726 xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, true);
1727
1728 if (VFS_I(ip)->i_nlink != 0) {
1729 /*
1730 * force is true because we are evicting an inode from the
1731 * cache. Post-eof blocks must be freed, lest we end up with
1732 * broken free space accounting.
1733 *
1734 * Note: don't bother with iolock here since lockdep complains
1735 * about acquiring it in reclaim context. We have the only
1736 * reference to the inode at this point anyways.
1737 */
1738 if (xfs_can_free_eofblocks(ip, true))
1739 xfs_free_eofblocks(ip);
1740
1741 goto out;
1742 }
1743
1744 if (S_ISREG(VFS_I(ip)->i_mode) &&
1745 (ip->i_disk_size != 0 || XFS_ISIZE(ip) != 0 ||
1746 ip->i_df.if_nextents > 0 || ip->i_delayed_blks > 0))
1747 truncate = 1;
1748
1749 error = xfs_qm_dqattach(ip);
1750 if (error)
1751 goto out;
1752
1753 if (S_ISLNK(VFS_I(ip)->i_mode))
1754 error = xfs_inactive_symlink(ip);
1755 else if (truncate)
1756 error = xfs_inactive_truncate(ip);
1757 if (error)
1758 goto out;
1759
1760 /*
1761 * If there are attributes associated with the file then blow them away
1762 * now. The code calls a routine that recursively deconstructs the
1763 * attribute fork. If also blows away the in-core attribute fork.
1764 */
1765 if (XFS_IFORK_Q(ip)) {
1766 error = xfs_attr_inactive(ip);
1767 if (error)
1768 goto out;
1769 }
1770
1771 ASSERT(!ip->i_afp);
1772 ASSERT(ip->i_forkoff == 0);
1773
1774 /*
1775 * Free the inode.
1776 */
1777 xfs_inactive_ifree(ip);
1778
1779 out:
1780 /*
1781 * We're done making metadata updates for this inode, so we can release
1782 * the attached dquots.
1783 */
1784 xfs_qm_dqdetach(ip);
1785 }
1786
1787 /*
1788 * In-Core Unlinked List Lookups
1789 * =============================
1790 *
1791 * Every inode is supposed to be reachable from some other piece of metadata
1792 * with the exception of the root directory. Inodes with a connection to a
1793 * file descriptor but not linked from anywhere in the on-disk directory tree
1794 * are collectively known as unlinked inodes, though the filesystem itself
1795 * maintains links to these inodes so that on-disk metadata are consistent.
1796 *
1797 * XFS implements a per-AG on-disk hash table of unlinked inodes. The AGI
1798 * header contains a number of buckets that point to an inode, and each inode
1799 * record has a pointer to the next inode in the hash chain. This
1800 * singly-linked list causes scaling problems in the iunlink remove function
1801 * because we must walk that list to find the inode that points to the inode
1802 * being removed from the unlinked hash bucket list.
1803 *
1804 * What if we modelled the unlinked list as a collection of records capturing
1805 * "X.next_unlinked = Y" relations? If we indexed those records on Y, we'd
1806 * have a fast way to look up unlinked list predecessors, which avoids the
1807 * slow list walk. That's exactly what we do here (in-core) with a per-AG
1808 * rhashtable.
1809 *
1810 * Because this is a backref cache, we ignore operational failures since the
1811 * iunlink code can fall back to the slow bucket walk. The only errors that
1812 * should bubble out are for obviously incorrect situations.
1813 *
1814 * All users of the backref cache MUST hold the AGI buffer lock to serialize
1815 * access or have otherwise provided for concurrency control.
1816 */
1817
1818 /* Capture a "X.next_unlinked = Y" relationship. */
1819 struct xfs_iunlink {
1820 struct rhash_head iu_rhash_head;
1821 xfs_agino_t iu_agino; /* X */
1822 xfs_agino_t iu_next_unlinked; /* Y */
1823 };
1824
1825 /* Unlinked list predecessor lookup hashtable construction */
1826 static int
xfs_iunlink_obj_cmpfn(struct rhashtable_compare_arg * arg,const void * obj)1827 xfs_iunlink_obj_cmpfn(
1828 struct rhashtable_compare_arg *arg,
1829 const void *obj)
1830 {
1831 const xfs_agino_t *key = arg->key;
1832 const struct xfs_iunlink *iu = obj;
1833
1834 if (iu->iu_next_unlinked != *key)
1835 return 1;
1836 return 0;
1837 }
1838
1839 static const struct rhashtable_params xfs_iunlink_hash_params = {
1840 .min_size = XFS_AGI_UNLINKED_BUCKETS,
1841 .key_len = sizeof(xfs_agino_t),
1842 .key_offset = offsetof(struct xfs_iunlink,
1843 iu_next_unlinked),
1844 .head_offset = offsetof(struct xfs_iunlink, iu_rhash_head),
1845 .automatic_shrinking = true,
1846 .obj_cmpfn = xfs_iunlink_obj_cmpfn,
1847 };
1848
1849 /*
1850 * Return X, where X.next_unlinked == @agino. Returns NULLAGINO if no such
1851 * relation is found.
1852 */
1853 static xfs_agino_t
xfs_iunlink_lookup_backref(struct xfs_perag * pag,xfs_agino_t agino)1854 xfs_iunlink_lookup_backref(
1855 struct xfs_perag *pag,
1856 xfs_agino_t agino)
1857 {
1858 struct xfs_iunlink *iu;
1859
1860 iu = rhashtable_lookup_fast(&pag->pagi_unlinked_hash, &agino,
1861 xfs_iunlink_hash_params);
1862 return iu ? iu->iu_agino : NULLAGINO;
1863 }
1864
1865 /*
1866 * Take ownership of an iunlink cache entry and insert it into the hash table.
1867 * If successful, the entry will be owned by the cache; if not, it is freed.
1868 * Either way, the caller does not own @iu after this call.
1869 */
1870 static int
xfs_iunlink_insert_backref(struct xfs_perag * pag,struct xfs_iunlink * iu)1871 xfs_iunlink_insert_backref(
1872 struct xfs_perag *pag,
1873 struct xfs_iunlink *iu)
1874 {
1875 int error;
1876
1877 error = rhashtable_insert_fast(&pag->pagi_unlinked_hash,
1878 &iu->iu_rhash_head, xfs_iunlink_hash_params);
1879 /*
1880 * Fail loudly if there already was an entry because that's a sign of
1881 * corruption of in-memory data. Also fail loudly if we see an error
1882 * code we didn't anticipate from the rhashtable code. Currently we
1883 * only anticipate ENOMEM.
1884 */
1885 if (error) {
1886 WARN(error != -ENOMEM, "iunlink cache insert error %d", error);
1887 kmem_free(iu);
1888 }
1889 /*
1890 * Absorb any runtime errors that aren't a result of corruption because
1891 * this is a cache and we can always fall back to bucket list scanning.
1892 */
1893 if (error != 0 && error != -EEXIST)
1894 error = 0;
1895 return error;
1896 }
1897
1898 /* Remember that @prev_agino.next_unlinked = @this_agino. */
1899 static int
xfs_iunlink_add_backref(struct xfs_perag * pag,xfs_agino_t prev_agino,xfs_agino_t this_agino)1900 xfs_iunlink_add_backref(
1901 struct xfs_perag *pag,
1902 xfs_agino_t prev_agino,
1903 xfs_agino_t this_agino)
1904 {
1905 struct xfs_iunlink *iu;
1906
1907 if (XFS_TEST_ERROR(false, pag->pag_mount, XFS_ERRTAG_IUNLINK_FALLBACK))
1908 return 0;
1909
1910 iu = kmem_zalloc(sizeof(*iu), KM_NOFS);
1911 iu->iu_agino = prev_agino;
1912 iu->iu_next_unlinked = this_agino;
1913
1914 return xfs_iunlink_insert_backref(pag, iu);
1915 }
1916
1917 /*
1918 * Replace X.next_unlinked = @agino with X.next_unlinked = @next_unlinked.
1919 * If @next_unlinked is NULLAGINO, we drop the backref and exit. If there
1920 * wasn't any such entry then we don't bother.
1921 */
1922 static int
xfs_iunlink_change_backref(struct xfs_perag * pag,xfs_agino_t agino,xfs_agino_t next_unlinked)1923 xfs_iunlink_change_backref(
1924 struct xfs_perag *pag,
1925 xfs_agino_t agino,
1926 xfs_agino_t next_unlinked)
1927 {
1928 struct xfs_iunlink *iu;
1929 int error;
1930
1931 /* Look up the old entry; if there wasn't one then exit. */
1932 iu = rhashtable_lookup_fast(&pag->pagi_unlinked_hash, &agino,
1933 xfs_iunlink_hash_params);
1934 if (!iu)
1935 return 0;
1936
1937 /*
1938 * Remove the entry. This shouldn't ever return an error, but if we
1939 * couldn't remove the old entry we don't want to add it again to the
1940 * hash table, and if the entry disappeared on us then someone's
1941 * violated the locking rules and we need to fail loudly. Either way
1942 * we cannot remove the inode because internal state is or would have
1943 * been corrupt.
1944 */
1945 error = rhashtable_remove_fast(&pag->pagi_unlinked_hash,
1946 &iu->iu_rhash_head, xfs_iunlink_hash_params);
1947 if (error)
1948 return error;
1949
1950 /* If there is no new next entry just free our item and return. */
1951 if (next_unlinked == NULLAGINO) {
1952 kmem_free(iu);
1953 return 0;
1954 }
1955
1956 /* Update the entry and re-add it to the hash table. */
1957 iu->iu_next_unlinked = next_unlinked;
1958 return xfs_iunlink_insert_backref(pag, iu);
1959 }
1960
1961 /* Set up the in-core predecessor structures. */
1962 int
xfs_iunlink_init(struct xfs_perag * pag)1963 xfs_iunlink_init(
1964 struct xfs_perag *pag)
1965 {
1966 return rhashtable_init(&pag->pagi_unlinked_hash,
1967 &xfs_iunlink_hash_params);
1968 }
1969
1970 /* Free the in-core predecessor structures. */
1971 static void
xfs_iunlink_free_item(void * ptr,void * arg)1972 xfs_iunlink_free_item(
1973 void *ptr,
1974 void *arg)
1975 {
1976 struct xfs_iunlink *iu = ptr;
1977 bool *freed_anything = arg;
1978
1979 *freed_anything = true;
1980 kmem_free(iu);
1981 }
1982
1983 void
xfs_iunlink_destroy(struct xfs_perag * pag)1984 xfs_iunlink_destroy(
1985 struct xfs_perag *pag)
1986 {
1987 bool freed_anything = false;
1988
1989 rhashtable_free_and_destroy(&pag->pagi_unlinked_hash,
1990 xfs_iunlink_free_item, &freed_anything);
1991
1992 ASSERT(freed_anything == false || xfs_is_shutdown(pag->pag_mount));
1993 }
1994
1995 /*
1996 * Point the AGI unlinked bucket at an inode and log the results. The caller
1997 * is responsible for validating the old value.
1998 */
1999 STATIC int
xfs_iunlink_update_bucket(struct xfs_trans * tp,struct xfs_perag * pag,struct xfs_buf * agibp,unsigned int bucket_index,xfs_agino_t new_agino)2000 xfs_iunlink_update_bucket(
2001 struct xfs_trans *tp,
2002 struct xfs_perag *pag,
2003 struct xfs_buf *agibp,
2004 unsigned int bucket_index,
2005 xfs_agino_t new_agino)
2006 {
2007 struct xfs_agi *agi = agibp->b_addr;
2008 xfs_agino_t old_value;
2009 int offset;
2010
2011 ASSERT(xfs_verify_agino_or_null(tp->t_mountp, pag->pag_agno, new_agino));
2012
2013 old_value = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2014 trace_xfs_iunlink_update_bucket(tp->t_mountp, pag->pag_agno, bucket_index,
2015 old_value, new_agino);
2016
2017 /*
2018 * We should never find the head of the list already set to the value
2019 * passed in because either we're adding or removing ourselves from the
2020 * head of the list.
2021 */
2022 if (old_value == new_agino) {
2023 xfs_buf_mark_corrupt(agibp);
2024 return -EFSCORRUPTED;
2025 }
2026
2027 agi->agi_unlinked[bucket_index] = cpu_to_be32(new_agino);
2028 offset = offsetof(struct xfs_agi, agi_unlinked) +
2029 (sizeof(xfs_agino_t) * bucket_index);
2030 xfs_trans_log_buf(tp, agibp, offset, offset + sizeof(xfs_agino_t) - 1);
2031 return 0;
2032 }
2033
2034 /* Set an on-disk inode's next_unlinked pointer. */
2035 STATIC void
xfs_iunlink_update_dinode(struct xfs_trans * tp,struct xfs_perag * pag,xfs_agino_t agino,struct xfs_buf * ibp,struct xfs_dinode * dip,struct xfs_imap * imap,xfs_agino_t next_agino)2036 xfs_iunlink_update_dinode(
2037 struct xfs_trans *tp,
2038 struct xfs_perag *pag,
2039 xfs_agino_t agino,
2040 struct xfs_buf *ibp,
2041 struct xfs_dinode *dip,
2042 struct xfs_imap *imap,
2043 xfs_agino_t next_agino)
2044 {
2045 struct xfs_mount *mp = tp->t_mountp;
2046 int offset;
2047
2048 ASSERT(xfs_verify_agino_or_null(mp, pag->pag_agno, next_agino));
2049
2050 trace_xfs_iunlink_update_dinode(mp, pag->pag_agno, agino,
2051 be32_to_cpu(dip->di_next_unlinked), next_agino);
2052
2053 dip->di_next_unlinked = cpu_to_be32(next_agino);
2054 offset = imap->im_boffset +
2055 offsetof(struct xfs_dinode, di_next_unlinked);
2056
2057 /* need to recalc the inode CRC if appropriate */
2058 xfs_dinode_calc_crc(mp, dip);
2059 xfs_trans_inode_buf(tp, ibp);
2060 xfs_trans_log_buf(tp, ibp, offset, offset + sizeof(xfs_agino_t) - 1);
2061 }
2062
2063 /* Set an in-core inode's unlinked pointer and return the old value. */
2064 STATIC int
xfs_iunlink_update_inode(struct xfs_trans * tp,struct xfs_inode * ip,struct xfs_perag * pag,xfs_agino_t next_agino,xfs_agino_t * old_next_agino)2065 xfs_iunlink_update_inode(
2066 struct xfs_trans *tp,
2067 struct xfs_inode *ip,
2068 struct xfs_perag *pag,
2069 xfs_agino_t next_agino,
2070 xfs_agino_t *old_next_agino)
2071 {
2072 struct xfs_mount *mp = tp->t_mountp;
2073 struct xfs_dinode *dip;
2074 struct xfs_buf *ibp;
2075 xfs_agino_t old_value;
2076 int error;
2077
2078 ASSERT(xfs_verify_agino_or_null(mp, pag->pag_agno, next_agino));
2079
2080 error = xfs_imap_to_bp(mp, tp, &ip->i_imap, &ibp);
2081 if (error)
2082 return error;
2083 dip = xfs_buf_offset(ibp, ip->i_imap.im_boffset);
2084
2085 /* Make sure the old pointer isn't garbage. */
2086 old_value = be32_to_cpu(dip->di_next_unlinked);
2087 if (!xfs_verify_agino_or_null(mp, pag->pag_agno, old_value)) {
2088 xfs_inode_verifier_error(ip, -EFSCORRUPTED, __func__, dip,
2089 sizeof(*dip), __this_address);
2090 error = -EFSCORRUPTED;
2091 goto out;
2092 }
2093
2094 /*
2095 * Since we're updating a linked list, we should never find that the
2096 * current pointer is the same as the new value, unless we're
2097 * terminating the list.
2098 */
2099 *old_next_agino = old_value;
2100 if (old_value == next_agino) {
2101 if (next_agino != NULLAGINO) {
2102 xfs_inode_verifier_error(ip, -EFSCORRUPTED, __func__,
2103 dip, sizeof(*dip), __this_address);
2104 error = -EFSCORRUPTED;
2105 }
2106 goto out;
2107 }
2108
2109 /* Ok, update the new pointer. */
2110 xfs_iunlink_update_dinode(tp, pag, XFS_INO_TO_AGINO(mp, ip->i_ino),
2111 ibp, dip, &ip->i_imap, next_agino);
2112 return 0;
2113 out:
2114 xfs_trans_brelse(tp, ibp);
2115 return error;
2116 }
2117
2118 /*
2119 * This is called when the inode's link count has gone to 0 or we are creating
2120 * a tmpfile via O_TMPFILE. The inode @ip must have nlink == 0.
2121 *
2122 * We place the on-disk inode on a list in the AGI. It will be pulled from this
2123 * list when the inode is freed.
2124 */
2125 STATIC int
xfs_iunlink(struct xfs_trans * tp,struct xfs_inode * ip)2126 xfs_iunlink(
2127 struct xfs_trans *tp,
2128 struct xfs_inode *ip)
2129 {
2130 struct xfs_mount *mp = tp->t_mountp;
2131 struct xfs_perag *pag;
2132 struct xfs_agi *agi;
2133 struct xfs_buf *agibp;
2134 xfs_agino_t next_agino;
2135 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2136 short bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
2137 int error;
2138
2139 ASSERT(VFS_I(ip)->i_nlink == 0);
2140 ASSERT(VFS_I(ip)->i_mode != 0);
2141 trace_xfs_iunlink(ip);
2142
2143 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
2144
2145 /* Get the agi buffer first. It ensures lock ordering on the list. */
2146 error = xfs_read_agi(mp, tp, pag->pag_agno, &agibp);
2147 if (error)
2148 goto out;
2149 agi = agibp->b_addr;
2150
2151 /*
2152 * Get the index into the agi hash table for the list this inode will
2153 * go on. Make sure the pointer isn't garbage and that this inode
2154 * isn't already on the list.
2155 */
2156 next_agino = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2157 if (next_agino == agino ||
2158 !xfs_verify_agino_or_null(mp, pag->pag_agno, next_agino)) {
2159 xfs_buf_mark_corrupt(agibp);
2160 error = -EFSCORRUPTED;
2161 goto out;
2162 }
2163
2164 if (next_agino != NULLAGINO) {
2165 xfs_agino_t old_agino;
2166
2167 /*
2168 * There is already another inode in the bucket, so point this
2169 * inode to the current head of the list.
2170 */
2171 error = xfs_iunlink_update_inode(tp, ip, pag, next_agino,
2172 &old_agino);
2173 if (error)
2174 goto out;
2175 ASSERT(old_agino == NULLAGINO);
2176
2177 /*
2178 * agino has been unlinked, add a backref from the next inode
2179 * back to agino.
2180 */
2181 error = xfs_iunlink_add_backref(pag, agino, next_agino);
2182 if (error)
2183 goto out;
2184 }
2185
2186 /* Point the head of the list to point to this inode. */
2187 error = xfs_iunlink_update_bucket(tp, pag, agibp, bucket_index, agino);
2188 out:
2189 xfs_perag_put(pag);
2190 return error;
2191 }
2192
2193 /* Return the imap, dinode pointer, and buffer for an inode. */
2194 STATIC int
xfs_iunlink_map_ino(struct xfs_trans * tp,xfs_agnumber_t agno,xfs_agino_t agino,struct xfs_imap * imap,struct xfs_dinode ** dipp,struct xfs_buf ** bpp)2195 xfs_iunlink_map_ino(
2196 struct xfs_trans *tp,
2197 xfs_agnumber_t agno,
2198 xfs_agino_t agino,
2199 struct xfs_imap *imap,
2200 struct xfs_dinode **dipp,
2201 struct xfs_buf **bpp)
2202 {
2203 struct xfs_mount *mp = tp->t_mountp;
2204 int error;
2205
2206 imap->im_blkno = 0;
2207 error = xfs_imap(mp, tp, XFS_AGINO_TO_INO(mp, agno, agino), imap, 0);
2208 if (error) {
2209 xfs_warn(mp, "%s: xfs_imap returned error %d.",
2210 __func__, error);
2211 return error;
2212 }
2213
2214 error = xfs_imap_to_bp(mp, tp, imap, bpp);
2215 if (error) {
2216 xfs_warn(mp, "%s: xfs_imap_to_bp returned error %d.",
2217 __func__, error);
2218 return error;
2219 }
2220
2221 *dipp = xfs_buf_offset(*bpp, imap->im_boffset);
2222 return 0;
2223 }
2224
2225 /*
2226 * Walk the unlinked chain from @head_agino until we find the inode that
2227 * points to @target_agino. Return the inode number, map, dinode pointer,
2228 * and inode cluster buffer of that inode as @agino, @imap, @dipp, and @bpp.
2229 *
2230 * @tp, @pag, @head_agino, and @target_agino are input parameters.
2231 * @agino, @imap, @dipp, and @bpp are all output parameters.
2232 *
2233 * Do not call this function if @target_agino is the head of the list.
2234 */
2235 STATIC int
xfs_iunlink_map_prev(struct xfs_trans * tp,struct xfs_perag * pag,xfs_agino_t head_agino,xfs_agino_t target_agino,xfs_agino_t * agino,struct xfs_imap * imap,struct xfs_dinode ** dipp,struct xfs_buf ** bpp)2236 xfs_iunlink_map_prev(
2237 struct xfs_trans *tp,
2238 struct xfs_perag *pag,
2239 xfs_agino_t head_agino,
2240 xfs_agino_t target_agino,
2241 xfs_agino_t *agino,
2242 struct xfs_imap *imap,
2243 struct xfs_dinode **dipp,
2244 struct xfs_buf **bpp)
2245 {
2246 struct xfs_mount *mp = tp->t_mountp;
2247 xfs_agino_t next_agino;
2248 int error;
2249
2250 ASSERT(head_agino != target_agino);
2251 *bpp = NULL;
2252
2253 /* See if our backref cache can find it faster. */
2254 *agino = xfs_iunlink_lookup_backref(pag, target_agino);
2255 if (*agino != NULLAGINO) {
2256 error = xfs_iunlink_map_ino(tp, pag->pag_agno, *agino, imap,
2257 dipp, bpp);
2258 if (error)
2259 return error;
2260
2261 if (be32_to_cpu((*dipp)->di_next_unlinked) == target_agino)
2262 return 0;
2263
2264 /*
2265 * If we get here the cache contents were corrupt, so drop the
2266 * buffer and fall back to walking the bucket list.
2267 */
2268 xfs_trans_brelse(tp, *bpp);
2269 *bpp = NULL;
2270 WARN_ON_ONCE(1);
2271 }
2272
2273 trace_xfs_iunlink_map_prev_fallback(mp, pag->pag_agno);
2274
2275 /* Otherwise, walk the entire bucket until we find it. */
2276 next_agino = head_agino;
2277 while (next_agino != target_agino) {
2278 xfs_agino_t unlinked_agino;
2279
2280 if (*bpp)
2281 xfs_trans_brelse(tp, *bpp);
2282
2283 *agino = next_agino;
2284 error = xfs_iunlink_map_ino(tp, pag->pag_agno, next_agino, imap,
2285 dipp, bpp);
2286 if (error)
2287 return error;
2288
2289 unlinked_agino = be32_to_cpu((*dipp)->di_next_unlinked);
2290 /*
2291 * Make sure this pointer is valid and isn't an obvious
2292 * infinite loop.
2293 */
2294 if (!xfs_verify_agino(mp, pag->pag_agno, unlinked_agino) ||
2295 next_agino == unlinked_agino) {
2296 XFS_CORRUPTION_ERROR(__func__,
2297 XFS_ERRLEVEL_LOW, mp,
2298 *dipp, sizeof(**dipp));
2299 error = -EFSCORRUPTED;
2300 return error;
2301 }
2302 next_agino = unlinked_agino;
2303 }
2304
2305 return 0;
2306 }
2307
2308 /*
2309 * Pull the on-disk inode from the AGI unlinked list.
2310 */
2311 STATIC int
xfs_iunlink_remove(struct xfs_trans * tp,struct xfs_perag * pag,struct xfs_inode * ip)2312 xfs_iunlink_remove(
2313 struct xfs_trans *tp,
2314 struct xfs_perag *pag,
2315 struct xfs_inode *ip)
2316 {
2317 struct xfs_mount *mp = tp->t_mountp;
2318 struct xfs_agi *agi;
2319 struct xfs_buf *agibp;
2320 struct xfs_buf *last_ibp;
2321 struct xfs_dinode *last_dip = NULL;
2322 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2323 xfs_agino_t next_agino;
2324 xfs_agino_t head_agino;
2325 short bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
2326 int error;
2327
2328 trace_xfs_iunlink_remove(ip);
2329
2330 /* Get the agi buffer first. It ensures lock ordering on the list. */
2331 error = xfs_read_agi(mp, tp, pag->pag_agno, &agibp);
2332 if (error)
2333 return error;
2334 agi = agibp->b_addr;
2335
2336 /*
2337 * Get the index into the agi hash table for the list this inode will
2338 * go on. Make sure the head pointer isn't garbage.
2339 */
2340 head_agino = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2341 if (!xfs_verify_agino(mp, pag->pag_agno, head_agino)) {
2342 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
2343 agi, sizeof(*agi));
2344 return -EFSCORRUPTED;
2345 }
2346
2347 /*
2348 * Set our inode's next_unlinked pointer to NULL and then return
2349 * the old pointer value so that we can update whatever was previous
2350 * to us in the list to point to whatever was next in the list.
2351 */
2352 error = xfs_iunlink_update_inode(tp, ip, pag, NULLAGINO, &next_agino);
2353 if (error)
2354 return error;
2355
2356 /*
2357 * If there was a backref pointing from the next inode back to this
2358 * one, remove it because we've removed this inode from the list.
2359 *
2360 * Later, if this inode was in the middle of the list we'll update
2361 * this inode's backref to point from the next inode.
2362 */
2363 if (next_agino != NULLAGINO) {
2364 error = xfs_iunlink_change_backref(pag, next_agino, NULLAGINO);
2365 if (error)
2366 return error;
2367 }
2368
2369 if (head_agino != agino) {
2370 struct xfs_imap imap;
2371 xfs_agino_t prev_agino;
2372
2373 /* We need to search the list for the inode being freed. */
2374 error = xfs_iunlink_map_prev(tp, pag, head_agino, agino,
2375 &prev_agino, &imap, &last_dip, &last_ibp);
2376 if (error)
2377 return error;
2378
2379 /* Point the previous inode on the list to the next inode. */
2380 xfs_iunlink_update_dinode(tp, pag, prev_agino, last_ibp,
2381 last_dip, &imap, next_agino);
2382
2383 /*
2384 * Now we deal with the backref for this inode. If this inode
2385 * pointed at a real inode, change the backref that pointed to
2386 * us to point to our old next. If this inode was the end of
2387 * the list, delete the backref that pointed to us. Note that
2388 * change_backref takes care of deleting the backref if
2389 * next_agino is NULLAGINO.
2390 */
2391 return xfs_iunlink_change_backref(agibp->b_pag, agino,
2392 next_agino);
2393 }
2394
2395 /* Point the head of the list to the next unlinked inode. */
2396 return xfs_iunlink_update_bucket(tp, pag, agibp, bucket_index,
2397 next_agino);
2398 }
2399
2400 /*
2401 * Look up the inode number specified and if it is not already marked XFS_ISTALE
2402 * mark it stale. We should only find clean inodes in this lookup that aren't
2403 * already stale.
2404 */
2405 static void
xfs_ifree_mark_inode_stale(struct xfs_perag * pag,struct xfs_inode * free_ip,xfs_ino_t inum)2406 xfs_ifree_mark_inode_stale(
2407 struct xfs_perag *pag,
2408 struct xfs_inode *free_ip,
2409 xfs_ino_t inum)
2410 {
2411 struct xfs_mount *mp = pag->pag_mount;
2412 struct xfs_inode_log_item *iip;
2413 struct xfs_inode *ip;
2414
2415 retry:
2416 rcu_read_lock();
2417 ip = radix_tree_lookup(&pag->pag_ici_root, XFS_INO_TO_AGINO(mp, inum));
2418
2419 /* Inode not in memory, nothing to do */
2420 if (!ip) {
2421 rcu_read_unlock();
2422 return;
2423 }
2424
2425 /*
2426 * because this is an RCU protected lookup, we could find a recently
2427 * freed or even reallocated inode during the lookup. We need to check
2428 * under the i_flags_lock for a valid inode here. Skip it if it is not
2429 * valid, the wrong inode or stale.
2430 */
2431 spin_lock(&ip->i_flags_lock);
2432 if (ip->i_ino != inum || __xfs_iflags_test(ip, XFS_ISTALE))
2433 goto out_iflags_unlock;
2434
2435 /*
2436 * Don't try to lock/unlock the current inode, but we _cannot_ skip the
2437 * other inodes that we did not find in the list attached to the buffer
2438 * and are not already marked stale. If we can't lock it, back off and
2439 * retry.
2440 */
2441 if (ip != free_ip) {
2442 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL)) {
2443 spin_unlock(&ip->i_flags_lock);
2444 rcu_read_unlock();
2445 delay(1);
2446 goto retry;
2447 }
2448 }
2449 ip->i_flags |= XFS_ISTALE;
2450
2451 /*
2452 * If the inode is flushing, it is already attached to the buffer. All
2453 * we needed to do here is mark the inode stale so buffer IO completion
2454 * will remove it from the AIL.
2455 */
2456 iip = ip->i_itemp;
2457 if (__xfs_iflags_test(ip, XFS_IFLUSHING)) {
2458 ASSERT(!list_empty(&iip->ili_item.li_bio_list));
2459 ASSERT(iip->ili_last_fields);
2460 goto out_iunlock;
2461 }
2462
2463 /*
2464 * Inodes not attached to the buffer can be released immediately.
2465 * Everything else has to go through xfs_iflush_abort() on journal
2466 * commit as the flock synchronises removal of the inode from the
2467 * cluster buffer against inode reclaim.
2468 */
2469 if (!iip || list_empty(&iip->ili_item.li_bio_list))
2470 goto out_iunlock;
2471
2472 __xfs_iflags_set(ip, XFS_IFLUSHING);
2473 spin_unlock(&ip->i_flags_lock);
2474 rcu_read_unlock();
2475
2476 /* we have a dirty inode in memory that has not yet been flushed. */
2477 spin_lock(&iip->ili_lock);
2478 iip->ili_last_fields = iip->ili_fields;
2479 iip->ili_fields = 0;
2480 iip->ili_fsync_fields = 0;
2481 spin_unlock(&iip->ili_lock);
2482 ASSERT(iip->ili_last_fields);
2483
2484 if (ip != free_ip)
2485 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2486 return;
2487
2488 out_iunlock:
2489 if (ip != free_ip)
2490 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2491 out_iflags_unlock:
2492 spin_unlock(&ip->i_flags_lock);
2493 rcu_read_unlock();
2494 }
2495
2496 /*
2497 * A big issue when freeing the inode cluster is that we _cannot_ skip any
2498 * inodes that are in memory - they all must be marked stale and attached to
2499 * the cluster buffer.
2500 */
2501 static int
xfs_ifree_cluster(struct xfs_trans * tp,struct xfs_perag * pag,struct xfs_inode * free_ip,struct xfs_icluster * xic)2502 xfs_ifree_cluster(
2503 struct xfs_trans *tp,
2504 struct xfs_perag *pag,
2505 struct xfs_inode *free_ip,
2506 struct xfs_icluster *xic)
2507 {
2508 struct xfs_mount *mp = free_ip->i_mount;
2509 struct xfs_ino_geometry *igeo = M_IGEO(mp);
2510 struct xfs_buf *bp;
2511 xfs_daddr_t blkno;
2512 xfs_ino_t inum = xic->first_ino;
2513 int nbufs;
2514 int i, j;
2515 int ioffset;
2516 int error;
2517
2518 nbufs = igeo->ialloc_blks / igeo->blocks_per_cluster;
2519
2520 for (j = 0; j < nbufs; j++, inum += igeo->inodes_per_cluster) {
2521 /*
2522 * The allocation bitmap tells us which inodes of the chunk were
2523 * physically allocated. Skip the cluster if an inode falls into
2524 * a sparse region.
2525 */
2526 ioffset = inum - xic->first_ino;
2527 if ((xic->alloc & XFS_INOBT_MASK(ioffset)) == 0) {
2528 ASSERT(ioffset % igeo->inodes_per_cluster == 0);
2529 continue;
2530 }
2531
2532 blkno = XFS_AGB_TO_DADDR(mp, XFS_INO_TO_AGNO(mp, inum),
2533 XFS_INO_TO_AGBNO(mp, inum));
2534
2535 /*
2536 * We obtain and lock the backing buffer first in the process
2537 * here to ensure dirty inodes attached to the buffer remain in
2538 * the flushing state while we mark them stale.
2539 *
2540 * If we scan the in-memory inodes first, then buffer IO can
2541 * complete before we get a lock on it, and hence we may fail
2542 * to mark all the active inodes on the buffer stale.
2543 */
2544 error = xfs_trans_get_buf(tp, mp->m_ddev_targp, blkno,
2545 mp->m_bsize * igeo->blocks_per_cluster,
2546 XBF_UNMAPPED, &bp);
2547 if (error)
2548 return error;
2549
2550 /*
2551 * This buffer may not have been correctly initialised as we
2552 * didn't read it from disk. That's not important because we are
2553 * only using to mark the buffer as stale in the log, and to
2554 * attach stale cached inodes on it. That means it will never be
2555 * dispatched for IO. If it is, we want to know about it, and we
2556 * want it to fail. We can acheive this by adding a write
2557 * verifier to the buffer.
2558 */
2559 bp->b_ops = &xfs_inode_buf_ops;
2560
2561 /*
2562 * Now we need to set all the cached clean inodes as XFS_ISTALE,
2563 * too. This requires lookups, and will skip inodes that we've
2564 * already marked XFS_ISTALE.
2565 */
2566 for (i = 0; i < igeo->inodes_per_cluster; i++)
2567 xfs_ifree_mark_inode_stale(pag, free_ip, inum + i);
2568
2569 xfs_trans_stale_inode_buf(tp, bp);
2570 xfs_trans_binval(tp, bp);
2571 }
2572 return 0;
2573 }
2574
2575 /*
2576 * This is called to return an inode to the inode free list. The inode should
2577 * already be truncated to 0 length and have no pages associated with it. This
2578 * routine also assumes that the inode is already a part of the transaction.
2579 *
2580 * The on-disk copy of the inode will have been added to the list of unlinked
2581 * inodes in the AGI. We need to remove the inode from that list atomically with
2582 * respect to freeing it here.
2583 */
2584 int
xfs_ifree(struct xfs_trans * tp,struct xfs_inode * ip)2585 xfs_ifree(
2586 struct xfs_trans *tp,
2587 struct xfs_inode *ip)
2588 {
2589 struct xfs_mount *mp = ip->i_mount;
2590 struct xfs_perag *pag;
2591 struct xfs_icluster xic = { 0 };
2592 struct xfs_inode_log_item *iip = ip->i_itemp;
2593 int error;
2594
2595 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
2596 ASSERT(VFS_I(ip)->i_nlink == 0);
2597 ASSERT(ip->i_df.if_nextents == 0);
2598 ASSERT(ip->i_disk_size == 0 || !S_ISREG(VFS_I(ip)->i_mode));
2599 ASSERT(ip->i_nblocks == 0);
2600
2601 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
2602
2603 /*
2604 * Free the inode first so that we guarantee that the AGI lock is going
2605 * to be taken before we remove the inode from the unlinked list. This
2606 * makes the AGI lock -> unlinked list modification order the same as
2607 * used in O_TMPFILE creation.
2608 */
2609 error = xfs_difree(tp, pag, ip->i_ino, &xic);
2610 if (error)
2611 goto out;
2612
2613 error = xfs_iunlink_remove(tp, pag, ip);
2614 if (error)
2615 goto out;
2616
2617 /*
2618 * Free any local-format data sitting around before we reset the
2619 * data fork to extents format. Note that the attr fork data has
2620 * already been freed by xfs_attr_inactive.
2621 */
2622 if (ip->i_df.if_format == XFS_DINODE_FMT_LOCAL) {
2623 kmem_free(ip->i_df.if_u1.if_data);
2624 ip->i_df.if_u1.if_data = NULL;
2625 ip->i_df.if_bytes = 0;
2626 }
2627
2628 VFS_I(ip)->i_mode = 0; /* mark incore inode as free */
2629 ip->i_diflags = 0;
2630 ip->i_diflags2 = mp->m_ino_geo.new_diflags2;
2631 ip->i_forkoff = 0; /* mark the attr fork not in use */
2632 ip->i_df.if_format = XFS_DINODE_FMT_EXTENTS;
2633 if (xfs_iflags_test(ip, XFS_IPRESERVE_DM_FIELDS))
2634 xfs_iflags_clear(ip, XFS_IPRESERVE_DM_FIELDS);
2635
2636 /* Don't attempt to replay owner changes for a deleted inode */
2637 spin_lock(&iip->ili_lock);
2638 iip->ili_fields &= ~(XFS_ILOG_AOWNER | XFS_ILOG_DOWNER);
2639 spin_unlock(&iip->ili_lock);
2640
2641 /*
2642 * Bump the generation count so no one will be confused
2643 * by reincarnations of this inode.
2644 */
2645 VFS_I(ip)->i_generation++;
2646 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2647
2648 if (xic.deleted)
2649 error = xfs_ifree_cluster(tp, pag, ip, &xic);
2650 out:
2651 xfs_perag_put(pag);
2652 return error;
2653 }
2654
2655 /*
2656 * This is called to unpin an inode. The caller must have the inode locked
2657 * in at least shared mode so that the buffer cannot be subsequently pinned
2658 * once someone is waiting for it to be unpinned.
2659 */
2660 static void
xfs_iunpin(struct xfs_inode * ip)2661 xfs_iunpin(
2662 struct xfs_inode *ip)
2663 {
2664 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
2665
2666 trace_xfs_inode_unpin_nowait(ip, _RET_IP_);
2667
2668 /* Give the log a push to start the unpinning I/O */
2669 xfs_log_force_seq(ip->i_mount, ip->i_itemp->ili_commit_seq, 0, NULL);
2670
2671 }
2672
2673 static void
__xfs_iunpin_wait(struct xfs_inode * ip)2674 __xfs_iunpin_wait(
2675 struct xfs_inode *ip)
2676 {
2677 wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_IPINNED_BIT);
2678 DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IPINNED_BIT);
2679
2680 xfs_iunpin(ip);
2681
2682 do {
2683 prepare_to_wait(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
2684 if (xfs_ipincount(ip))
2685 io_schedule();
2686 } while (xfs_ipincount(ip));
2687 finish_wait(wq, &wait.wq_entry);
2688 }
2689
2690 void
xfs_iunpin_wait(struct xfs_inode * ip)2691 xfs_iunpin_wait(
2692 struct xfs_inode *ip)
2693 {
2694 if (xfs_ipincount(ip))
2695 __xfs_iunpin_wait(ip);
2696 }
2697
2698 /*
2699 * Removing an inode from the namespace involves removing the directory entry
2700 * and dropping the link count on the inode. Removing the directory entry can
2701 * result in locking an AGF (directory blocks were freed) and removing a link
2702 * count can result in placing the inode on an unlinked list which results in
2703 * locking an AGI.
2704 *
2705 * The big problem here is that we have an ordering constraint on AGF and AGI
2706 * locking - inode allocation locks the AGI, then can allocate a new extent for
2707 * new inodes, locking the AGF after the AGI. Similarly, freeing the inode
2708 * removes the inode from the unlinked list, requiring that we lock the AGI
2709 * first, and then freeing the inode can result in an inode chunk being freed
2710 * and hence freeing disk space requiring that we lock an AGF.
2711 *
2712 * Hence the ordering that is imposed by other parts of the code is AGI before
2713 * AGF. This means we cannot remove the directory entry before we drop the inode
2714 * reference count and put it on the unlinked list as this results in a lock
2715 * order of AGF then AGI, and this can deadlock against inode allocation and
2716 * freeing. Therefore we must drop the link counts before we remove the
2717 * directory entry.
2718 *
2719 * This is still safe from a transactional point of view - it is not until we
2720 * get to xfs_defer_finish() that we have the possibility of multiple
2721 * transactions in this operation. Hence as long as we remove the directory
2722 * entry and drop the link count in the first transaction of the remove
2723 * operation, there are no transactional constraints on the ordering here.
2724 */
2725 int
xfs_remove(xfs_inode_t * dp,struct xfs_name * name,xfs_inode_t * ip)2726 xfs_remove(
2727 xfs_inode_t *dp,
2728 struct xfs_name *name,
2729 xfs_inode_t *ip)
2730 {
2731 xfs_mount_t *mp = dp->i_mount;
2732 xfs_trans_t *tp = NULL;
2733 int is_dir = S_ISDIR(VFS_I(ip)->i_mode);
2734 int dontcare;
2735 int error = 0;
2736 uint resblks;
2737
2738 trace_xfs_remove(dp, name);
2739
2740 if (xfs_is_shutdown(mp))
2741 return -EIO;
2742
2743 error = xfs_qm_dqattach(dp);
2744 if (error)
2745 goto std_return;
2746
2747 error = xfs_qm_dqattach(ip);
2748 if (error)
2749 goto std_return;
2750
2751 /*
2752 * We try to get the real space reservation first, allowing for
2753 * directory btree deletion(s) implying possible bmap insert(s). If we
2754 * can't get the space reservation then we use 0 instead, and avoid the
2755 * bmap btree insert(s) in the directory code by, if the bmap insert
2756 * tries to happen, instead trimming the LAST block from the directory.
2757 *
2758 * Ignore EDQUOT and ENOSPC being returned via nospace_error because
2759 * the directory code can handle a reservationless update and we don't
2760 * want to prevent a user from trying to free space by deleting things.
2761 */
2762 resblks = XFS_REMOVE_SPACE_RES(mp);
2763 error = xfs_trans_alloc_dir(dp, &M_RES(mp)->tr_remove, ip, &resblks,
2764 &tp, &dontcare);
2765 if (error) {
2766 ASSERT(error != -ENOSPC);
2767 goto std_return;
2768 }
2769
2770 /*
2771 * If we're removing a directory perform some additional validation.
2772 */
2773 if (is_dir) {
2774 ASSERT(VFS_I(ip)->i_nlink >= 2);
2775 if (VFS_I(ip)->i_nlink != 2) {
2776 error = -ENOTEMPTY;
2777 goto out_trans_cancel;
2778 }
2779 if (!xfs_dir_isempty(ip)) {
2780 error = -ENOTEMPTY;
2781 goto out_trans_cancel;
2782 }
2783
2784 /* Drop the link from ip's "..". */
2785 error = xfs_droplink(tp, dp);
2786 if (error)
2787 goto out_trans_cancel;
2788
2789 /* Drop the "." link from ip to self. */
2790 error = xfs_droplink(tp, ip);
2791 if (error)
2792 goto out_trans_cancel;
2793
2794 /*
2795 * Point the unlinked child directory's ".." entry to the root
2796 * directory to eliminate back-references to inodes that may
2797 * get freed before the child directory is closed. If the fs
2798 * gets shrunk, this can lead to dirent inode validation errors.
2799 */
2800 if (dp->i_ino != tp->t_mountp->m_sb.sb_rootino) {
2801 error = xfs_dir_replace(tp, ip, &xfs_name_dotdot,
2802 tp->t_mountp->m_sb.sb_rootino, 0);
2803 if (error)
2804 return error;
2805 }
2806 } else {
2807 /*
2808 * When removing a non-directory we need to log the parent
2809 * inode here. For a directory this is done implicitly
2810 * by the xfs_droplink call for the ".." entry.
2811 */
2812 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
2813 }
2814 xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2815
2816 /* Drop the link from dp to ip. */
2817 error = xfs_droplink(tp, ip);
2818 if (error)
2819 goto out_trans_cancel;
2820
2821 error = xfs_dir_removename(tp, dp, name, ip->i_ino, resblks);
2822 if (error) {
2823 ASSERT(error != -ENOENT);
2824 goto out_trans_cancel;
2825 }
2826
2827 /*
2828 * If this is a synchronous mount, make sure that the
2829 * remove transaction goes to disk before returning to
2830 * the user.
2831 */
2832 if (xfs_has_wsync(mp) || xfs_has_dirsync(mp))
2833 xfs_trans_set_sync(tp);
2834
2835 error = xfs_trans_commit(tp);
2836 if (error)
2837 goto std_return;
2838
2839 if (is_dir && xfs_inode_is_filestream(ip))
2840 xfs_filestream_deassociate(ip);
2841
2842 return 0;
2843
2844 out_trans_cancel:
2845 xfs_trans_cancel(tp);
2846 std_return:
2847 return error;
2848 }
2849
2850 /*
2851 * Enter all inodes for a rename transaction into a sorted array.
2852 */
2853 #define __XFS_SORT_INODES 5
2854 STATIC void
xfs_sort_for_rename(struct xfs_inode * dp1,struct xfs_inode * dp2,struct xfs_inode * ip1,struct xfs_inode * ip2,struct xfs_inode * wip,struct xfs_inode ** i_tab,int * num_inodes)2855 xfs_sort_for_rename(
2856 struct xfs_inode *dp1, /* in: old (source) directory inode */
2857 struct xfs_inode *dp2, /* in: new (target) directory inode */
2858 struct xfs_inode *ip1, /* in: inode of old entry */
2859 struct xfs_inode *ip2, /* in: inode of new entry */
2860 struct xfs_inode *wip, /* in: whiteout inode */
2861 struct xfs_inode **i_tab,/* out: sorted array of inodes */
2862 int *num_inodes) /* in/out: inodes in array */
2863 {
2864 int i, j;
2865
2866 ASSERT(*num_inodes == __XFS_SORT_INODES);
2867 memset(i_tab, 0, *num_inodes * sizeof(struct xfs_inode *));
2868
2869 /*
2870 * i_tab contains a list of pointers to inodes. We initialize
2871 * the table here & we'll sort it. We will then use it to
2872 * order the acquisition of the inode locks.
2873 *
2874 * Note that the table may contain duplicates. e.g., dp1 == dp2.
2875 */
2876 i = 0;
2877 i_tab[i++] = dp1;
2878 i_tab[i++] = dp2;
2879 i_tab[i++] = ip1;
2880 if (ip2)
2881 i_tab[i++] = ip2;
2882 if (wip)
2883 i_tab[i++] = wip;
2884 *num_inodes = i;
2885
2886 /*
2887 * Sort the elements via bubble sort. (Remember, there are at
2888 * most 5 elements to sort, so this is adequate.)
2889 */
2890 for (i = 0; i < *num_inodes; i++) {
2891 for (j = 1; j < *num_inodes; j++) {
2892 if (i_tab[j]->i_ino < i_tab[j-1]->i_ino) {
2893 struct xfs_inode *temp = i_tab[j];
2894 i_tab[j] = i_tab[j-1];
2895 i_tab[j-1] = temp;
2896 }
2897 }
2898 }
2899 }
2900
2901 static int
xfs_finish_rename(struct xfs_trans * tp)2902 xfs_finish_rename(
2903 struct xfs_trans *tp)
2904 {
2905 /*
2906 * If this is a synchronous mount, make sure that the rename transaction
2907 * goes to disk before returning to the user.
2908 */
2909 if (xfs_has_wsync(tp->t_mountp) || xfs_has_dirsync(tp->t_mountp))
2910 xfs_trans_set_sync(tp);
2911
2912 return xfs_trans_commit(tp);
2913 }
2914
2915 /*
2916 * xfs_cross_rename()
2917 *
2918 * responsible for handling RENAME_EXCHANGE flag in renameat2() syscall
2919 */
2920 STATIC int
xfs_cross_rename(struct xfs_trans * tp,struct xfs_inode * dp1,struct xfs_name * name1,struct xfs_inode * ip1,struct xfs_inode * dp2,struct xfs_name * name2,struct xfs_inode * ip2,int spaceres)2921 xfs_cross_rename(
2922 struct xfs_trans *tp,
2923 struct xfs_inode *dp1,
2924 struct xfs_name *name1,
2925 struct xfs_inode *ip1,
2926 struct xfs_inode *dp2,
2927 struct xfs_name *name2,
2928 struct xfs_inode *ip2,
2929 int spaceres)
2930 {
2931 int error = 0;
2932 int ip1_flags = 0;
2933 int ip2_flags = 0;
2934 int dp2_flags = 0;
2935
2936 /* Swap inode number for dirent in first parent */
2937 error = xfs_dir_replace(tp, dp1, name1, ip2->i_ino, spaceres);
2938 if (error)
2939 goto out_trans_abort;
2940
2941 /* Swap inode number for dirent in second parent */
2942 error = xfs_dir_replace(tp, dp2, name2, ip1->i_ino, spaceres);
2943 if (error)
2944 goto out_trans_abort;
2945
2946 /*
2947 * If we're renaming one or more directories across different parents,
2948 * update the respective ".." entries (and link counts) to match the new
2949 * parents.
2950 */
2951 if (dp1 != dp2) {
2952 dp2_flags = XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2953
2954 if (S_ISDIR(VFS_I(ip2)->i_mode)) {
2955 error = xfs_dir_replace(tp, ip2, &xfs_name_dotdot,
2956 dp1->i_ino, spaceres);
2957 if (error)
2958 goto out_trans_abort;
2959
2960 /* transfer ip2 ".." reference to dp1 */
2961 if (!S_ISDIR(VFS_I(ip1)->i_mode)) {
2962 error = xfs_droplink(tp, dp2);
2963 if (error)
2964 goto out_trans_abort;
2965 xfs_bumplink(tp, dp1);
2966 }
2967
2968 /*
2969 * Although ip1 isn't changed here, userspace needs
2970 * to be warned about the change, so that applications
2971 * relying on it (like backup ones), will properly
2972 * notify the change
2973 */
2974 ip1_flags |= XFS_ICHGTIME_CHG;
2975 ip2_flags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2976 }
2977
2978 if (S_ISDIR(VFS_I(ip1)->i_mode)) {
2979 error = xfs_dir_replace(tp, ip1, &xfs_name_dotdot,
2980 dp2->i_ino, spaceres);
2981 if (error)
2982 goto out_trans_abort;
2983
2984 /* transfer ip1 ".." reference to dp2 */
2985 if (!S_ISDIR(VFS_I(ip2)->i_mode)) {
2986 error = xfs_droplink(tp, dp1);
2987 if (error)
2988 goto out_trans_abort;
2989 xfs_bumplink(tp, dp2);
2990 }
2991
2992 /*
2993 * Although ip2 isn't changed here, userspace needs
2994 * to be warned about the change, so that applications
2995 * relying on it (like backup ones), will properly
2996 * notify the change
2997 */
2998 ip1_flags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2999 ip2_flags |= XFS_ICHGTIME_CHG;
3000 }
3001 }
3002
3003 if (ip1_flags) {
3004 xfs_trans_ichgtime(tp, ip1, ip1_flags);
3005 xfs_trans_log_inode(tp, ip1, XFS_ILOG_CORE);
3006 }
3007 if (ip2_flags) {
3008 xfs_trans_ichgtime(tp, ip2, ip2_flags);
3009 xfs_trans_log_inode(tp, ip2, XFS_ILOG_CORE);
3010 }
3011 if (dp2_flags) {
3012 xfs_trans_ichgtime(tp, dp2, dp2_flags);
3013 xfs_trans_log_inode(tp, dp2, XFS_ILOG_CORE);
3014 }
3015 xfs_trans_ichgtime(tp, dp1, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3016 xfs_trans_log_inode(tp, dp1, XFS_ILOG_CORE);
3017 return xfs_finish_rename(tp);
3018
3019 out_trans_abort:
3020 xfs_trans_cancel(tp);
3021 return error;
3022 }
3023
3024 /*
3025 * xfs_rename_alloc_whiteout()
3026 *
3027 * Return a referenced, unlinked, unlocked inode that can be used as a
3028 * whiteout in a rename transaction. We use a tmpfile inode here so that if we
3029 * crash between allocating the inode and linking it into the rename transaction
3030 * recovery will free the inode and we won't leak it.
3031 */
3032 static int
xfs_rename_alloc_whiteout(struct user_namespace * mnt_userns,struct xfs_inode * dp,struct xfs_inode ** wip)3033 xfs_rename_alloc_whiteout(
3034 struct user_namespace *mnt_userns,
3035 struct xfs_inode *dp,
3036 struct xfs_inode **wip)
3037 {
3038 struct xfs_inode *tmpfile;
3039 int error;
3040
3041 error = xfs_create_tmpfile(mnt_userns, dp, S_IFCHR | WHITEOUT_MODE,
3042 &tmpfile);
3043 if (error)
3044 return error;
3045
3046 /*
3047 * Prepare the tmpfile inode as if it were created through the VFS.
3048 * Complete the inode setup and flag it as linkable. nlink is already
3049 * zero, so we can skip the drop_nlink.
3050 */
3051 xfs_setup_iops(tmpfile);
3052 xfs_finish_inode_setup(tmpfile);
3053 VFS_I(tmpfile)->i_state |= I_LINKABLE;
3054
3055 *wip = tmpfile;
3056 return 0;
3057 }
3058
3059 /*
3060 * xfs_rename
3061 */
3062 int
xfs_rename(struct user_namespace * mnt_userns,struct xfs_inode * src_dp,struct xfs_name * src_name,struct xfs_inode * src_ip,struct xfs_inode * target_dp,struct xfs_name * target_name,struct xfs_inode * target_ip,unsigned int flags)3063 xfs_rename(
3064 struct user_namespace *mnt_userns,
3065 struct xfs_inode *src_dp,
3066 struct xfs_name *src_name,
3067 struct xfs_inode *src_ip,
3068 struct xfs_inode *target_dp,
3069 struct xfs_name *target_name,
3070 struct xfs_inode *target_ip,
3071 unsigned int flags)
3072 {
3073 struct xfs_mount *mp = src_dp->i_mount;
3074 struct xfs_trans *tp;
3075 struct xfs_inode *wip = NULL; /* whiteout inode */
3076 struct xfs_inode *inodes[__XFS_SORT_INODES];
3077 int i;
3078 int num_inodes = __XFS_SORT_INODES;
3079 bool new_parent = (src_dp != target_dp);
3080 bool src_is_directory = S_ISDIR(VFS_I(src_ip)->i_mode);
3081 int spaceres;
3082 bool retried = false;
3083 int error, nospace_error = 0;
3084
3085 trace_xfs_rename(src_dp, target_dp, src_name, target_name);
3086
3087 if ((flags & RENAME_EXCHANGE) && !target_ip)
3088 return -EINVAL;
3089
3090 /*
3091 * If we are doing a whiteout operation, allocate the whiteout inode
3092 * we will be placing at the target and ensure the type is set
3093 * appropriately.
3094 */
3095 if (flags & RENAME_WHITEOUT) {
3096 error = xfs_rename_alloc_whiteout(mnt_userns, target_dp, &wip);
3097 if (error)
3098 return error;
3099
3100 /* setup target dirent info as whiteout */
3101 src_name->type = XFS_DIR3_FT_CHRDEV;
3102 }
3103
3104 xfs_sort_for_rename(src_dp, target_dp, src_ip, target_ip, wip,
3105 inodes, &num_inodes);
3106
3107 retry:
3108 nospace_error = 0;
3109 spaceres = XFS_RENAME_SPACE_RES(mp, target_name->len);
3110 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, spaceres, 0, 0, &tp);
3111 if (error == -ENOSPC) {
3112 nospace_error = error;
3113 spaceres = 0;
3114 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, 0, 0, 0,
3115 &tp);
3116 }
3117 if (error)
3118 goto out_release_wip;
3119
3120 /*
3121 * Attach the dquots to the inodes
3122 */
3123 error = xfs_qm_vop_rename_dqattach(inodes);
3124 if (error)
3125 goto out_trans_cancel;
3126
3127 /*
3128 * Lock all the participating inodes. Depending upon whether
3129 * the target_name exists in the target directory, and
3130 * whether the target directory is the same as the source
3131 * directory, we can lock from 2 to 4 inodes.
3132 */
3133 xfs_lock_inodes(inodes, num_inodes, XFS_ILOCK_EXCL);
3134
3135 /*
3136 * Join all the inodes to the transaction. From this point on,
3137 * we can rely on either trans_commit or trans_cancel to unlock
3138 * them.
3139 */
3140 xfs_trans_ijoin(tp, src_dp, XFS_ILOCK_EXCL);
3141 if (new_parent)
3142 xfs_trans_ijoin(tp, target_dp, XFS_ILOCK_EXCL);
3143 xfs_trans_ijoin(tp, src_ip, XFS_ILOCK_EXCL);
3144 if (target_ip)
3145 xfs_trans_ijoin(tp, target_ip, XFS_ILOCK_EXCL);
3146 if (wip)
3147 xfs_trans_ijoin(tp, wip, XFS_ILOCK_EXCL);
3148
3149 /*
3150 * If we are using project inheritance, we only allow renames
3151 * into our tree when the project IDs are the same; else the
3152 * tree quota mechanism would be circumvented.
3153 */
3154 if (unlikely((target_dp->i_diflags & XFS_DIFLAG_PROJINHERIT) &&
3155 target_dp->i_projid != src_ip->i_projid)) {
3156 error = -EXDEV;
3157 goto out_trans_cancel;
3158 }
3159
3160 /* RENAME_EXCHANGE is unique from here on. */
3161 if (flags & RENAME_EXCHANGE)
3162 return xfs_cross_rename(tp, src_dp, src_name, src_ip,
3163 target_dp, target_name, target_ip,
3164 spaceres);
3165
3166 /*
3167 * Try to reserve quota to handle an expansion of the target directory.
3168 * We'll allow the rename to continue in reservationless mode if we hit
3169 * a space usage constraint. If we trigger reservationless mode, save
3170 * the errno if there isn't any free space in the target directory.
3171 */
3172 if (spaceres != 0) {
3173 error = xfs_trans_reserve_quota_nblks(tp, target_dp, spaceres,
3174 0, false);
3175 if (error == -EDQUOT || error == -ENOSPC) {
3176 if (!retried) {
3177 xfs_trans_cancel(tp);
3178 xfs_blockgc_free_quota(target_dp, 0);
3179 retried = true;
3180 goto retry;
3181 }
3182
3183 nospace_error = error;
3184 spaceres = 0;
3185 error = 0;
3186 }
3187 if (error)
3188 goto out_trans_cancel;
3189 }
3190
3191 /*
3192 * Check for expected errors before we dirty the transaction
3193 * so we can return an error without a transaction abort.
3194 */
3195 if (target_ip == NULL) {
3196 /*
3197 * If there's no space reservation, check the entry will
3198 * fit before actually inserting it.
3199 */
3200 if (!spaceres) {
3201 error = xfs_dir_canenter(tp, target_dp, target_name);
3202 if (error)
3203 goto out_trans_cancel;
3204 }
3205 } else {
3206 /*
3207 * If target exists and it's a directory, check that whether
3208 * it can be destroyed.
3209 */
3210 if (S_ISDIR(VFS_I(target_ip)->i_mode) &&
3211 (!xfs_dir_isempty(target_ip) ||
3212 (VFS_I(target_ip)->i_nlink > 2))) {
3213 error = -EEXIST;
3214 goto out_trans_cancel;
3215 }
3216 }
3217
3218 /*
3219 * Lock the AGI buffers we need to handle bumping the nlink of the
3220 * whiteout inode off the unlinked list and to handle dropping the
3221 * nlink of the target inode. Per locking order rules, do this in
3222 * increasing AG order and before directory block allocation tries to
3223 * grab AGFs because we grab AGIs before AGFs.
3224 *
3225 * The (vfs) caller must ensure that if src is a directory then
3226 * target_ip is either null or an empty directory.
3227 */
3228 for (i = 0; i < num_inodes && inodes[i] != NULL; i++) {
3229 if (inodes[i] == wip ||
3230 (inodes[i] == target_ip &&
3231 (VFS_I(target_ip)->i_nlink == 1 || src_is_directory))) {
3232 struct xfs_buf *bp;
3233 xfs_agnumber_t agno;
3234
3235 agno = XFS_INO_TO_AGNO(mp, inodes[i]->i_ino);
3236 error = xfs_read_agi(mp, tp, agno, &bp);
3237 if (error)
3238 goto out_trans_cancel;
3239 }
3240 }
3241
3242 /*
3243 * Directory entry creation below may acquire the AGF. Remove
3244 * the whiteout from the unlinked list first to preserve correct
3245 * AGI/AGF locking order. This dirties the transaction so failures
3246 * after this point will abort and log recovery will clean up the
3247 * mess.
3248 *
3249 * For whiteouts, we need to bump the link count on the whiteout
3250 * inode. After this point, we have a real link, clear the tmpfile
3251 * state flag from the inode so it doesn't accidentally get misused
3252 * in future.
3253 */
3254 if (wip) {
3255 struct xfs_perag *pag;
3256
3257 ASSERT(VFS_I(wip)->i_nlink == 0);
3258
3259 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, wip->i_ino));
3260 error = xfs_iunlink_remove(tp, pag, wip);
3261 xfs_perag_put(pag);
3262 if (error)
3263 goto out_trans_cancel;
3264
3265 xfs_bumplink(tp, wip);
3266 VFS_I(wip)->i_state &= ~I_LINKABLE;
3267 }
3268
3269 /*
3270 * Set up the target.
3271 */
3272 if (target_ip == NULL) {
3273 /*
3274 * If target does not exist and the rename crosses
3275 * directories, adjust the target directory link count
3276 * to account for the ".." reference from the new entry.
3277 */
3278 error = xfs_dir_createname(tp, target_dp, target_name,
3279 src_ip->i_ino, spaceres);
3280 if (error)
3281 goto out_trans_cancel;
3282
3283 xfs_trans_ichgtime(tp, target_dp,
3284 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3285
3286 if (new_parent && src_is_directory) {
3287 xfs_bumplink(tp, target_dp);
3288 }
3289 } else { /* target_ip != NULL */
3290 /*
3291 * Link the source inode under the target name.
3292 * If the source inode is a directory and we are moving
3293 * it across directories, its ".." entry will be
3294 * inconsistent until we replace that down below.
3295 *
3296 * In case there is already an entry with the same
3297 * name at the destination directory, remove it first.
3298 */
3299 error = xfs_dir_replace(tp, target_dp, target_name,
3300 src_ip->i_ino, spaceres);
3301 if (error)
3302 goto out_trans_cancel;
3303
3304 xfs_trans_ichgtime(tp, target_dp,
3305 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3306
3307 /*
3308 * Decrement the link count on the target since the target
3309 * dir no longer points to it.
3310 */
3311 error = xfs_droplink(tp, target_ip);
3312 if (error)
3313 goto out_trans_cancel;
3314
3315 if (src_is_directory) {
3316 /*
3317 * Drop the link from the old "." entry.
3318 */
3319 error = xfs_droplink(tp, target_ip);
3320 if (error)
3321 goto out_trans_cancel;
3322 }
3323 } /* target_ip != NULL */
3324
3325 /*
3326 * Remove the source.
3327 */
3328 if (new_parent && src_is_directory) {
3329 /*
3330 * Rewrite the ".." entry to point to the new
3331 * directory.
3332 */
3333 error = xfs_dir_replace(tp, src_ip, &xfs_name_dotdot,
3334 target_dp->i_ino, spaceres);
3335 ASSERT(error != -EEXIST);
3336 if (error)
3337 goto out_trans_cancel;
3338 }
3339
3340 /*
3341 * We always want to hit the ctime on the source inode.
3342 *
3343 * This isn't strictly required by the standards since the source
3344 * inode isn't really being changed, but old unix file systems did
3345 * it and some incremental backup programs won't work without it.
3346 */
3347 xfs_trans_ichgtime(tp, src_ip, XFS_ICHGTIME_CHG);
3348 xfs_trans_log_inode(tp, src_ip, XFS_ILOG_CORE);
3349
3350 /*
3351 * Adjust the link count on src_dp. This is necessary when
3352 * renaming a directory, either within one parent when
3353 * the target existed, or across two parent directories.
3354 */
3355 if (src_is_directory && (new_parent || target_ip != NULL)) {
3356
3357 /*
3358 * Decrement link count on src_directory since the
3359 * entry that's moved no longer points to it.
3360 */
3361 error = xfs_droplink(tp, src_dp);
3362 if (error)
3363 goto out_trans_cancel;
3364 }
3365
3366 /*
3367 * For whiteouts, we only need to update the source dirent with the
3368 * inode number of the whiteout inode rather than removing it
3369 * altogether.
3370 */
3371 if (wip)
3372 error = xfs_dir_replace(tp, src_dp, src_name, wip->i_ino,
3373 spaceres);
3374 else
3375 error = xfs_dir_removename(tp, src_dp, src_name, src_ip->i_ino,
3376 spaceres);
3377
3378 if (error)
3379 goto out_trans_cancel;
3380
3381 xfs_trans_ichgtime(tp, src_dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3382 xfs_trans_log_inode(tp, src_dp, XFS_ILOG_CORE);
3383 if (new_parent)
3384 xfs_trans_log_inode(tp, target_dp, XFS_ILOG_CORE);
3385
3386 error = xfs_finish_rename(tp);
3387 if (wip)
3388 xfs_irele(wip);
3389 return error;
3390
3391 out_trans_cancel:
3392 xfs_trans_cancel(tp);
3393 out_release_wip:
3394 if (wip)
3395 xfs_irele(wip);
3396 if (error == -ENOSPC && nospace_error)
3397 error = nospace_error;
3398 return error;
3399 }
3400
3401 static int
xfs_iflush(struct xfs_inode * ip,struct xfs_buf * bp)3402 xfs_iflush(
3403 struct xfs_inode *ip,
3404 struct xfs_buf *bp)
3405 {
3406 struct xfs_inode_log_item *iip = ip->i_itemp;
3407 struct xfs_dinode *dip;
3408 struct xfs_mount *mp = ip->i_mount;
3409 int error;
3410
3411 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
3412 ASSERT(xfs_iflags_test(ip, XFS_IFLUSHING));
3413 ASSERT(ip->i_df.if_format != XFS_DINODE_FMT_BTREE ||
3414 ip->i_df.if_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
3415 ASSERT(iip->ili_item.li_buf == bp);
3416
3417 dip = xfs_buf_offset(bp, ip->i_imap.im_boffset);
3418
3419 /*
3420 * We don't flush the inode if any of the following checks fail, but we
3421 * do still update the log item and attach to the backing buffer as if
3422 * the flush happened. This is a formality to facilitate predictable
3423 * error handling as the caller will shutdown and fail the buffer.
3424 */
3425 error = -EFSCORRUPTED;
3426 if (XFS_TEST_ERROR(dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC),
3427 mp, XFS_ERRTAG_IFLUSH_1)) {
3428 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3429 "%s: Bad inode %Lu magic number 0x%x, ptr "PTR_FMT,
3430 __func__, ip->i_ino, be16_to_cpu(dip->di_magic), dip);
3431 goto flush_out;
3432 }
3433 if (S_ISREG(VFS_I(ip)->i_mode)) {
3434 if (XFS_TEST_ERROR(
3435 ip->i_df.if_format != XFS_DINODE_FMT_EXTENTS &&
3436 ip->i_df.if_format != XFS_DINODE_FMT_BTREE,
3437 mp, XFS_ERRTAG_IFLUSH_3)) {
3438 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3439 "%s: Bad regular inode %Lu, ptr "PTR_FMT,
3440 __func__, ip->i_ino, ip);
3441 goto flush_out;
3442 }
3443 } else if (S_ISDIR(VFS_I(ip)->i_mode)) {
3444 if (XFS_TEST_ERROR(
3445 ip->i_df.if_format != XFS_DINODE_FMT_EXTENTS &&
3446 ip->i_df.if_format != XFS_DINODE_FMT_BTREE &&
3447 ip->i_df.if_format != XFS_DINODE_FMT_LOCAL,
3448 mp, XFS_ERRTAG_IFLUSH_4)) {
3449 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3450 "%s: Bad directory inode %Lu, ptr "PTR_FMT,
3451 __func__, ip->i_ino, ip);
3452 goto flush_out;
3453 }
3454 }
3455 if (XFS_TEST_ERROR(ip->i_df.if_nextents + xfs_ifork_nextents(ip->i_afp) >
3456 ip->i_nblocks, mp, XFS_ERRTAG_IFLUSH_5)) {
3457 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3458 "%s: detected corrupt incore inode %llu, "
3459 "total extents = %llu nblocks = %lld, ptr "PTR_FMT,
3460 __func__, ip->i_ino,
3461 ip->i_df.if_nextents + xfs_ifork_nextents(ip->i_afp),
3462 ip->i_nblocks, ip);
3463 goto flush_out;
3464 }
3465 if (XFS_TEST_ERROR(ip->i_forkoff > mp->m_sb.sb_inodesize,
3466 mp, XFS_ERRTAG_IFLUSH_6)) {
3467 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3468 "%s: bad inode %Lu, forkoff 0x%x, ptr "PTR_FMT,
3469 __func__, ip->i_ino, ip->i_forkoff, ip);
3470 goto flush_out;
3471 }
3472
3473 /*
3474 * Inode item log recovery for v2 inodes are dependent on the flushiter
3475 * count for correct sequencing. We bump the flush iteration count so
3476 * we can detect flushes which postdate a log record during recovery.
3477 * This is redundant as we now log every change and hence this can't
3478 * happen but we need to still do it to ensure backwards compatibility
3479 * with old kernels that predate logging all inode changes.
3480 */
3481 if (!xfs_has_v3inodes(mp))
3482 ip->i_flushiter++;
3483
3484 /*
3485 * If there are inline format data / attr forks attached to this inode,
3486 * make sure they are not corrupt.
3487 */
3488 if (ip->i_df.if_format == XFS_DINODE_FMT_LOCAL &&
3489 xfs_ifork_verify_local_data(ip))
3490 goto flush_out;
3491 if (ip->i_afp && ip->i_afp->if_format == XFS_DINODE_FMT_LOCAL &&
3492 xfs_ifork_verify_local_attr(ip))
3493 goto flush_out;
3494
3495 /*
3496 * Copy the dirty parts of the inode into the on-disk inode. We always
3497 * copy out the core of the inode, because if the inode is dirty at all
3498 * the core must be.
3499 */
3500 xfs_inode_to_disk(ip, dip, iip->ili_item.li_lsn);
3501
3502 /* Wrap, we never let the log put out DI_MAX_FLUSH */
3503 if (!xfs_has_v3inodes(mp)) {
3504 if (ip->i_flushiter == DI_MAX_FLUSH)
3505 ip->i_flushiter = 0;
3506 }
3507
3508 xfs_iflush_fork(ip, dip, iip, XFS_DATA_FORK);
3509 if (XFS_IFORK_Q(ip))
3510 xfs_iflush_fork(ip, dip, iip, XFS_ATTR_FORK);
3511
3512 /*
3513 * We've recorded everything logged in the inode, so we'd like to clear
3514 * the ili_fields bits so we don't log and flush things unnecessarily.
3515 * However, we can't stop logging all this information until the data
3516 * we've copied into the disk buffer is written to disk. If we did we
3517 * might overwrite the copy of the inode in the log with all the data
3518 * after re-logging only part of it, and in the face of a crash we
3519 * wouldn't have all the data we need to recover.
3520 *
3521 * What we do is move the bits to the ili_last_fields field. When
3522 * logging the inode, these bits are moved back to the ili_fields field.
3523 * In the xfs_buf_inode_iodone() routine we clear ili_last_fields, since
3524 * we know that the information those bits represent is permanently on
3525 * disk. As long as the flush completes before the inode is logged
3526 * again, then both ili_fields and ili_last_fields will be cleared.
3527 */
3528 error = 0;
3529 flush_out:
3530 spin_lock(&iip->ili_lock);
3531 iip->ili_last_fields = iip->ili_fields;
3532 iip->ili_fields = 0;
3533 iip->ili_fsync_fields = 0;
3534 spin_unlock(&iip->ili_lock);
3535
3536 /*
3537 * Store the current LSN of the inode so that we can tell whether the
3538 * item has moved in the AIL from xfs_buf_inode_iodone().
3539 */
3540 xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
3541 &iip->ili_item.li_lsn);
3542
3543 /* generate the checksum. */
3544 xfs_dinode_calc_crc(mp, dip);
3545 return error;
3546 }
3547
3548 /*
3549 * Non-blocking flush of dirty inode metadata into the backing buffer.
3550 *
3551 * The caller must have a reference to the inode and hold the cluster buffer
3552 * locked. The function will walk across all the inodes on the cluster buffer it
3553 * can find and lock without blocking, and flush them to the cluster buffer.
3554 *
3555 * On successful flushing of at least one inode, the caller must write out the
3556 * buffer and release it. If no inodes are flushed, -EAGAIN will be returned and
3557 * the caller needs to release the buffer. On failure, the filesystem will be
3558 * shut down, the buffer will have been unlocked and released, and EFSCORRUPTED
3559 * will be returned.
3560 */
3561 int
xfs_iflush_cluster(struct xfs_buf * bp)3562 xfs_iflush_cluster(
3563 struct xfs_buf *bp)
3564 {
3565 struct xfs_mount *mp = bp->b_mount;
3566 struct xfs_log_item *lip, *n;
3567 struct xfs_inode *ip;
3568 struct xfs_inode_log_item *iip;
3569 int clcount = 0;
3570 int error = 0;
3571
3572 /*
3573 * We must use the safe variant here as on shutdown xfs_iflush_abort()
3574 * will remove itself from the list.
3575 */
3576 list_for_each_entry_safe(lip, n, &bp->b_li_list, li_bio_list) {
3577 iip = (struct xfs_inode_log_item *)lip;
3578 ip = iip->ili_inode;
3579
3580 /*
3581 * Quick and dirty check to avoid locks if possible.
3582 */
3583 if (__xfs_iflags_test(ip, XFS_IRECLAIM | XFS_IFLUSHING))
3584 continue;
3585 if (xfs_ipincount(ip))
3586 continue;
3587
3588 /*
3589 * The inode is still attached to the buffer, which means it is
3590 * dirty but reclaim might try to grab it. Check carefully for
3591 * that, and grab the ilock while still holding the i_flags_lock
3592 * to guarantee reclaim will not be able to reclaim this inode
3593 * once we drop the i_flags_lock.
3594 */
3595 spin_lock(&ip->i_flags_lock);
3596 ASSERT(!__xfs_iflags_test(ip, XFS_ISTALE));
3597 if (__xfs_iflags_test(ip, XFS_IRECLAIM | XFS_IFLUSHING)) {
3598 spin_unlock(&ip->i_flags_lock);
3599 continue;
3600 }
3601
3602 /*
3603 * ILOCK will pin the inode against reclaim and prevent
3604 * concurrent transactions modifying the inode while we are
3605 * flushing the inode. If we get the lock, set the flushing
3606 * state before we drop the i_flags_lock.
3607 */
3608 if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
3609 spin_unlock(&ip->i_flags_lock);
3610 continue;
3611 }
3612 __xfs_iflags_set(ip, XFS_IFLUSHING);
3613 spin_unlock(&ip->i_flags_lock);
3614
3615 /*
3616 * Abort flushing this inode if we are shut down because the
3617 * inode may not currently be in the AIL. This can occur when
3618 * log I/O failure unpins the inode without inserting into the
3619 * AIL, leaving a dirty/unpinned inode attached to the buffer
3620 * that otherwise looks like it should be flushed.
3621 */
3622 if (xlog_is_shutdown(mp->m_log)) {
3623 xfs_iunpin_wait(ip);
3624 xfs_iflush_abort(ip);
3625 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3626 error = -EIO;
3627 continue;
3628 }
3629
3630 /* don't block waiting on a log force to unpin dirty inodes */
3631 if (xfs_ipincount(ip)) {
3632 xfs_iflags_clear(ip, XFS_IFLUSHING);
3633 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3634 continue;
3635 }
3636
3637 if (!xfs_inode_clean(ip))
3638 error = xfs_iflush(ip, bp);
3639 else
3640 xfs_iflags_clear(ip, XFS_IFLUSHING);
3641 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3642 if (error)
3643 break;
3644 clcount++;
3645 }
3646
3647 if (error) {
3648 /*
3649 * Shutdown first so we kill the log before we release this
3650 * buffer. If it is an INODE_ALLOC buffer and pins the tail
3651 * of the log, failing it before the _log_ is shut down can
3652 * result in the log tail being moved forward in the journal
3653 * on disk because log writes can still be taking place. Hence
3654 * unpinning the tail will allow the ICREATE intent to be
3655 * removed from the log an recovery will fail with uninitialised
3656 * inode cluster buffers.
3657 */
3658 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
3659 bp->b_flags |= XBF_ASYNC;
3660 xfs_buf_ioend_fail(bp);
3661 return error;
3662 }
3663
3664 if (!clcount)
3665 return -EAGAIN;
3666
3667 XFS_STATS_INC(mp, xs_icluster_flushcnt);
3668 XFS_STATS_ADD(mp, xs_icluster_flushinode, clcount);
3669 return 0;
3670
3671 }
3672
3673 /* Release an inode. */
3674 void
xfs_irele(struct xfs_inode * ip)3675 xfs_irele(
3676 struct xfs_inode *ip)
3677 {
3678 trace_xfs_irele(ip, _RET_IP_);
3679 iput(VFS_I(ip));
3680 }
3681
3682 /*
3683 * Ensure all commited transactions touching the inode are written to the log.
3684 */
3685 int
xfs_log_force_inode(struct xfs_inode * ip)3686 xfs_log_force_inode(
3687 struct xfs_inode *ip)
3688 {
3689 xfs_csn_t seq = 0;
3690
3691 xfs_ilock(ip, XFS_ILOCK_SHARED);
3692 if (xfs_ipincount(ip))
3693 seq = ip->i_itemp->ili_commit_seq;
3694 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3695
3696 if (!seq)
3697 return 0;
3698 return xfs_log_force_seq(ip->i_mount, seq, XFS_LOG_SYNC, NULL);
3699 }
3700
3701 /*
3702 * Grab the exclusive iolock for a data copy from src to dest, making sure to
3703 * abide vfs locking order (lowest pointer value goes first) and breaking the
3704 * layout leases before proceeding. The loop is needed because we cannot call
3705 * the blocking break_layout() with the iolocks held, and therefore have to
3706 * back out both locks.
3707 */
3708 static int
xfs_iolock_two_inodes_and_break_layout(struct inode * src,struct inode * dest)3709 xfs_iolock_two_inodes_and_break_layout(
3710 struct inode *src,
3711 struct inode *dest)
3712 {
3713 int error;
3714
3715 if (src > dest)
3716 swap(src, dest);
3717
3718 retry:
3719 /* Wait to break both inodes' layouts before we start locking. */
3720 error = break_layout(src, true);
3721 if (error)
3722 return error;
3723 if (src != dest) {
3724 error = break_layout(dest, true);
3725 if (error)
3726 return error;
3727 }
3728
3729 /* Lock one inode and make sure nobody got in and leased it. */
3730 inode_lock(src);
3731 error = break_layout(src, false);
3732 if (error) {
3733 inode_unlock(src);
3734 if (error == -EWOULDBLOCK)
3735 goto retry;
3736 return error;
3737 }
3738
3739 if (src == dest)
3740 return 0;
3741
3742 /* Lock the other inode and make sure nobody got in and leased it. */
3743 inode_lock_nested(dest, I_MUTEX_NONDIR2);
3744 error = break_layout(dest, false);
3745 if (error) {
3746 inode_unlock(src);
3747 inode_unlock(dest);
3748 if (error == -EWOULDBLOCK)
3749 goto retry;
3750 return error;
3751 }
3752
3753 return 0;
3754 }
3755
3756 /*
3757 * Lock two inodes so that userspace cannot initiate I/O via file syscalls or
3758 * mmap activity.
3759 */
3760 int
xfs_ilock2_io_mmap(struct xfs_inode * ip1,struct xfs_inode * ip2)3761 xfs_ilock2_io_mmap(
3762 struct xfs_inode *ip1,
3763 struct xfs_inode *ip2)
3764 {
3765 int ret;
3766
3767 ret = xfs_iolock_two_inodes_and_break_layout(VFS_I(ip1), VFS_I(ip2));
3768 if (ret)
3769 return ret;
3770 filemap_invalidate_lock_two(VFS_I(ip1)->i_mapping,
3771 VFS_I(ip2)->i_mapping);
3772 return 0;
3773 }
3774
3775 /* Unlock both inodes to allow IO and mmap activity. */
3776 void
xfs_iunlock2_io_mmap(struct xfs_inode * ip1,struct xfs_inode * ip2)3777 xfs_iunlock2_io_mmap(
3778 struct xfs_inode *ip1,
3779 struct xfs_inode *ip2)
3780 {
3781 filemap_invalidate_unlock_two(VFS_I(ip1)->i_mapping,
3782 VFS_I(ip2)->i_mapping);
3783 inode_unlock(VFS_I(ip2));
3784 if (ip1 != ip2)
3785 inode_unlock(VFS_I(ip1));
3786 }
3787