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_iunlink_item.h"
24 #include "xfs_ialloc.h"
25 #include "xfs_bmap.h"
26 #include "xfs_bmap_util.h"
27 #include "xfs_errortag.h"
28 #include "xfs_error.h"
29 #include "xfs_quota.h"
30 #include "xfs_filestream.h"
31 #include "xfs_trace.h"
32 #include "xfs_icache.h"
33 #include "xfs_symlink.h"
34 #include "xfs_trans_priv.h"
35 #include "xfs_log.h"
36 #include "xfs_bmap_btree.h"
37 #include "xfs_reflink.h"
38 #include "xfs_ag.h"
39 #include "xfs_log_priv.h"
40
41 struct kmem_cache *xfs_inode_cache;
42
43 /*
44 * Used in xfs_itruncate_extents(). This is the maximum number of extents
45 * freed from a file in a single transaction.
46 */
47 #define XFS_ITRUNC_MAX_EXTENTS 2
48
49 STATIC int xfs_iunlink(struct xfs_trans *, struct xfs_inode *);
50 STATIC int xfs_iunlink_remove(struct xfs_trans *tp, struct xfs_perag *pag,
51 struct xfs_inode *);
52
53 /*
54 * helper function to extract extent size hint from inode
55 */
56 xfs_extlen_t
xfs_get_extsz_hint(struct xfs_inode * ip)57 xfs_get_extsz_hint(
58 struct xfs_inode *ip)
59 {
60 /*
61 * No point in aligning allocations if we need to COW to actually
62 * write to them.
63 */
64 if (xfs_is_always_cow_inode(ip))
65 return 0;
66 if ((ip->i_diflags & XFS_DIFLAG_EXTSIZE) && ip->i_extsize)
67 return ip->i_extsize;
68 if (XFS_IS_REALTIME_INODE(ip))
69 return ip->i_mount->m_sb.sb_rextsize;
70 return 0;
71 }
72
73 /*
74 * Helper function to extract CoW extent size hint from inode.
75 * Between the extent size hint and the CoW extent size hint, we
76 * return the greater of the two. If the value is zero (automatic),
77 * use the default size.
78 */
79 xfs_extlen_t
xfs_get_cowextsz_hint(struct xfs_inode * ip)80 xfs_get_cowextsz_hint(
81 struct xfs_inode *ip)
82 {
83 xfs_extlen_t a, b;
84
85 a = 0;
86 if (ip->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE)
87 a = ip->i_cowextsize;
88 b = xfs_get_extsz_hint(ip);
89
90 a = max(a, b);
91 if (a == 0)
92 return XFS_DEFAULT_COWEXTSZ_HINT;
93 return a;
94 }
95
96 /*
97 * These two are wrapper routines around the xfs_ilock() routine used to
98 * centralize some grungy code. They are used in places that wish to lock the
99 * inode solely for reading the extents. The reason these places can't just
100 * call xfs_ilock(ip, XFS_ILOCK_SHARED) is that the inode lock also guards to
101 * bringing in of the extents from disk for a file in b-tree format. If the
102 * inode is in b-tree format, then we need to lock the inode exclusively until
103 * the extents are read in. Locking it exclusively all the time would limit
104 * our parallelism unnecessarily, though. What we do instead is check to see
105 * if the extents have been read in yet, and only lock the inode exclusively
106 * if they have not.
107 *
108 * The functions return a value which should be given to the corresponding
109 * xfs_iunlock() call.
110 */
111 uint
xfs_ilock_data_map_shared(struct xfs_inode * ip)112 xfs_ilock_data_map_shared(
113 struct xfs_inode *ip)
114 {
115 uint lock_mode = XFS_ILOCK_SHARED;
116
117 if (xfs_need_iread_extents(&ip->i_df))
118 lock_mode = XFS_ILOCK_EXCL;
119 xfs_ilock(ip, lock_mode);
120 return lock_mode;
121 }
122
123 uint
xfs_ilock_attr_map_shared(struct xfs_inode * ip)124 xfs_ilock_attr_map_shared(
125 struct xfs_inode *ip)
126 {
127 uint lock_mode = XFS_ILOCK_SHARED;
128
129 if (xfs_inode_has_attr_fork(ip) && xfs_need_iread_extents(&ip->i_af))
130 lock_mode = XFS_ILOCK_EXCL;
131 xfs_ilock(ip, lock_mode);
132 return lock_mode;
133 }
134
135 /*
136 * You can't set both SHARED and EXCL for the same lock,
137 * and only XFS_IOLOCK_SHARED, XFS_IOLOCK_EXCL, XFS_MMAPLOCK_SHARED,
138 * XFS_MMAPLOCK_EXCL, XFS_ILOCK_SHARED, XFS_ILOCK_EXCL are valid values
139 * to set in lock_flags.
140 */
141 static inline void
xfs_lock_flags_assert(uint lock_flags)142 xfs_lock_flags_assert(
143 uint lock_flags)
144 {
145 ASSERT((lock_flags & (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL)) !=
146 (XFS_IOLOCK_SHARED | XFS_IOLOCK_EXCL));
147 ASSERT((lock_flags & (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL)) !=
148 (XFS_MMAPLOCK_SHARED | XFS_MMAPLOCK_EXCL));
149 ASSERT((lock_flags & (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL)) !=
150 (XFS_ILOCK_SHARED | XFS_ILOCK_EXCL));
151 ASSERT((lock_flags & ~(XFS_LOCK_MASK | XFS_LOCK_SUBCLASS_MASK)) == 0);
152 ASSERT(lock_flags != 0);
153 }
154
155 /*
156 * In addition to i_rwsem in the VFS inode, the xfs inode contains 2
157 * multi-reader locks: invalidate_lock and the i_lock. This routine allows
158 * various combinations of the locks to be obtained.
159 *
160 * The 3 locks should always be ordered so that the IO lock is obtained first,
161 * the mmap lock second and the ilock last in order to prevent deadlock.
162 *
163 * Basic locking order:
164 *
165 * i_rwsem -> invalidate_lock -> page_lock -> i_ilock
166 *
167 * mmap_lock locking order:
168 *
169 * i_rwsem -> page lock -> mmap_lock
170 * mmap_lock -> invalidate_lock -> page_lock
171 *
172 * The difference in mmap_lock locking order mean that we cannot hold the
173 * invalidate_lock over syscall based read(2)/write(2) based IO. These IO paths
174 * can fault in pages during copy in/out (for buffered IO) or require the
175 * mmap_lock in get_user_pages() to map the user pages into the kernel address
176 * space for direct IO. Similarly the i_rwsem cannot be taken inside a page
177 * fault because page faults already hold the mmap_lock.
178 *
179 * Hence to serialise fully against both syscall and mmap based IO, we need to
180 * take both the i_rwsem and the invalidate_lock. These locks should *only* be
181 * both taken in places where we need to invalidate the page cache in a race
182 * free manner (e.g. truncate, hole punch and other extent manipulation
183 * functions).
184 */
185 void
xfs_ilock(xfs_inode_t * ip,uint lock_flags)186 xfs_ilock(
187 xfs_inode_t *ip,
188 uint lock_flags)
189 {
190 trace_xfs_ilock(ip, lock_flags, _RET_IP_);
191
192 xfs_lock_flags_assert(lock_flags);
193
194 if (lock_flags & XFS_IOLOCK_EXCL) {
195 down_write_nested(&VFS_I(ip)->i_rwsem,
196 XFS_IOLOCK_DEP(lock_flags));
197 } else if (lock_flags & XFS_IOLOCK_SHARED) {
198 down_read_nested(&VFS_I(ip)->i_rwsem,
199 XFS_IOLOCK_DEP(lock_flags));
200 }
201
202 if (lock_flags & XFS_MMAPLOCK_EXCL) {
203 down_write_nested(&VFS_I(ip)->i_mapping->invalidate_lock,
204 XFS_MMAPLOCK_DEP(lock_flags));
205 } else if (lock_flags & XFS_MMAPLOCK_SHARED) {
206 down_read_nested(&VFS_I(ip)->i_mapping->invalidate_lock,
207 XFS_MMAPLOCK_DEP(lock_flags));
208 }
209
210 if (lock_flags & XFS_ILOCK_EXCL)
211 mrupdate_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
212 else if (lock_flags & XFS_ILOCK_SHARED)
213 mraccess_nested(&ip->i_lock, XFS_ILOCK_DEP(lock_flags));
214 }
215
216 /*
217 * This is just like xfs_ilock(), except that the caller
218 * is guaranteed not to sleep. It returns 1 if it gets
219 * the requested locks and 0 otherwise. If the IO lock is
220 * obtained but the inode lock cannot be, then the IO lock
221 * is dropped before returning.
222 *
223 * ip -- the inode being locked
224 * lock_flags -- this parameter indicates the inode's locks to be
225 * to be locked. See the comment for xfs_ilock() for a list
226 * of valid values.
227 */
228 int
xfs_ilock_nowait(xfs_inode_t * ip,uint lock_flags)229 xfs_ilock_nowait(
230 xfs_inode_t *ip,
231 uint lock_flags)
232 {
233 trace_xfs_ilock_nowait(ip, lock_flags, _RET_IP_);
234
235 xfs_lock_flags_assert(lock_flags);
236
237 if (lock_flags & XFS_IOLOCK_EXCL) {
238 if (!down_write_trylock(&VFS_I(ip)->i_rwsem))
239 goto out;
240 } else if (lock_flags & XFS_IOLOCK_SHARED) {
241 if (!down_read_trylock(&VFS_I(ip)->i_rwsem))
242 goto out;
243 }
244
245 if (lock_flags & XFS_MMAPLOCK_EXCL) {
246 if (!down_write_trylock(&VFS_I(ip)->i_mapping->invalidate_lock))
247 goto out_undo_iolock;
248 } else if (lock_flags & XFS_MMAPLOCK_SHARED) {
249 if (!down_read_trylock(&VFS_I(ip)->i_mapping->invalidate_lock))
250 goto out_undo_iolock;
251 }
252
253 if (lock_flags & XFS_ILOCK_EXCL) {
254 if (!mrtryupdate(&ip->i_lock))
255 goto out_undo_mmaplock;
256 } else if (lock_flags & XFS_ILOCK_SHARED) {
257 if (!mrtryaccess(&ip->i_lock))
258 goto out_undo_mmaplock;
259 }
260 return 1;
261
262 out_undo_mmaplock:
263 if (lock_flags & XFS_MMAPLOCK_EXCL)
264 up_write(&VFS_I(ip)->i_mapping->invalidate_lock);
265 else if (lock_flags & XFS_MMAPLOCK_SHARED)
266 up_read(&VFS_I(ip)->i_mapping->invalidate_lock);
267 out_undo_iolock:
268 if (lock_flags & XFS_IOLOCK_EXCL)
269 up_write(&VFS_I(ip)->i_rwsem);
270 else if (lock_flags & XFS_IOLOCK_SHARED)
271 up_read(&VFS_I(ip)->i_rwsem);
272 out:
273 return 0;
274 }
275
276 /*
277 * xfs_iunlock() is used to drop the inode locks acquired with
278 * xfs_ilock() and xfs_ilock_nowait(). The caller must pass
279 * in the flags given to xfs_ilock() or xfs_ilock_nowait() so
280 * that we know which locks to drop.
281 *
282 * ip -- the inode being unlocked
283 * lock_flags -- this parameter indicates the inode's locks to be
284 * to be unlocked. See the comment for xfs_ilock() for a list
285 * of valid values for this parameter.
286 *
287 */
288 void
xfs_iunlock(xfs_inode_t * ip,uint lock_flags)289 xfs_iunlock(
290 xfs_inode_t *ip,
291 uint lock_flags)
292 {
293 xfs_lock_flags_assert(lock_flags);
294
295 if (lock_flags & XFS_IOLOCK_EXCL)
296 up_write(&VFS_I(ip)->i_rwsem);
297 else if (lock_flags & XFS_IOLOCK_SHARED)
298 up_read(&VFS_I(ip)->i_rwsem);
299
300 if (lock_flags & XFS_MMAPLOCK_EXCL)
301 up_write(&VFS_I(ip)->i_mapping->invalidate_lock);
302 else if (lock_flags & XFS_MMAPLOCK_SHARED)
303 up_read(&VFS_I(ip)->i_mapping->invalidate_lock);
304
305 if (lock_flags & XFS_ILOCK_EXCL)
306 mrunlock_excl(&ip->i_lock);
307 else if (lock_flags & XFS_ILOCK_SHARED)
308 mrunlock_shared(&ip->i_lock);
309
310 trace_xfs_iunlock(ip, lock_flags, _RET_IP_);
311 }
312
313 /*
314 * give up write locks. the i/o lock cannot be held nested
315 * if it is being demoted.
316 */
317 void
xfs_ilock_demote(xfs_inode_t * ip,uint lock_flags)318 xfs_ilock_demote(
319 xfs_inode_t *ip,
320 uint lock_flags)
321 {
322 ASSERT(lock_flags & (XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL));
323 ASSERT((lock_flags &
324 ~(XFS_IOLOCK_EXCL|XFS_MMAPLOCK_EXCL|XFS_ILOCK_EXCL)) == 0);
325
326 if (lock_flags & XFS_ILOCK_EXCL)
327 mrdemote(&ip->i_lock);
328 if (lock_flags & XFS_MMAPLOCK_EXCL)
329 downgrade_write(&VFS_I(ip)->i_mapping->invalidate_lock);
330 if (lock_flags & XFS_IOLOCK_EXCL)
331 downgrade_write(&VFS_I(ip)->i_rwsem);
332
333 trace_xfs_ilock_demote(ip, lock_flags, _RET_IP_);
334 }
335
336 #if defined(DEBUG) || defined(XFS_WARN)
337 static inline bool
__xfs_rwsem_islocked(struct rw_semaphore * rwsem,bool shared)338 __xfs_rwsem_islocked(
339 struct rw_semaphore *rwsem,
340 bool shared)
341 {
342 if (!debug_locks)
343 return rwsem_is_locked(rwsem);
344
345 if (!shared)
346 return lockdep_is_held_type(rwsem, 0);
347
348 /*
349 * We are checking that the lock is held at least in shared
350 * mode but don't care that it might be held exclusively
351 * (i.e. shared | excl). Hence we check if the lock is held
352 * in any mode rather than an explicit shared mode.
353 */
354 return lockdep_is_held_type(rwsem, -1);
355 }
356
357 bool
xfs_isilocked(struct xfs_inode * ip,uint lock_flags)358 xfs_isilocked(
359 struct xfs_inode *ip,
360 uint lock_flags)
361 {
362 if (lock_flags & (XFS_ILOCK_EXCL|XFS_ILOCK_SHARED)) {
363 if (!(lock_flags & XFS_ILOCK_SHARED))
364 return !!ip->i_lock.mr_writer;
365 return rwsem_is_locked(&ip->i_lock.mr_lock);
366 }
367
368 if (lock_flags & (XFS_MMAPLOCK_EXCL|XFS_MMAPLOCK_SHARED)) {
369 return __xfs_rwsem_islocked(&VFS_I(ip)->i_mapping->invalidate_lock,
370 (lock_flags & XFS_MMAPLOCK_SHARED));
371 }
372
373 if (lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) {
374 return __xfs_rwsem_islocked(&VFS_I(ip)->i_rwsem,
375 (lock_flags & XFS_IOLOCK_SHARED));
376 }
377
378 ASSERT(0);
379 return false;
380 }
381 #endif
382
383 /*
384 * xfs_lockdep_subclass_ok() is only used in an ASSERT, so is only called when
385 * DEBUG or XFS_WARN is set. And MAX_LOCKDEP_SUBCLASSES is then only defined
386 * when CONFIG_LOCKDEP is set. Hence the complex define below to avoid build
387 * errors and warnings.
388 */
389 #if (defined(DEBUG) || defined(XFS_WARN)) && defined(CONFIG_LOCKDEP)
390 static bool
xfs_lockdep_subclass_ok(int subclass)391 xfs_lockdep_subclass_ok(
392 int subclass)
393 {
394 return subclass < MAX_LOCKDEP_SUBCLASSES;
395 }
396 #else
397 #define xfs_lockdep_subclass_ok(subclass) (true)
398 #endif
399
400 /*
401 * Bump the subclass so xfs_lock_inodes() acquires each lock with a different
402 * value. This can be called for any type of inode lock combination, including
403 * parent locking. Care must be taken to ensure we don't overrun the subclass
404 * storage fields in the class mask we build.
405 */
406 static inline uint
xfs_lock_inumorder(uint lock_mode,uint subclass)407 xfs_lock_inumorder(
408 uint lock_mode,
409 uint subclass)
410 {
411 uint class = 0;
412
413 ASSERT(!(lock_mode & (XFS_ILOCK_PARENT | XFS_ILOCK_RTBITMAP |
414 XFS_ILOCK_RTSUM)));
415 ASSERT(xfs_lockdep_subclass_ok(subclass));
416
417 if (lock_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)) {
418 ASSERT(subclass <= XFS_IOLOCK_MAX_SUBCLASS);
419 class += subclass << XFS_IOLOCK_SHIFT;
420 }
421
422 if (lock_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)) {
423 ASSERT(subclass <= XFS_MMAPLOCK_MAX_SUBCLASS);
424 class += subclass << XFS_MMAPLOCK_SHIFT;
425 }
426
427 if (lock_mode & (XFS_ILOCK_SHARED|XFS_ILOCK_EXCL)) {
428 ASSERT(subclass <= XFS_ILOCK_MAX_SUBCLASS);
429 class += subclass << XFS_ILOCK_SHIFT;
430 }
431
432 return (lock_mode & ~XFS_LOCK_SUBCLASS_MASK) | class;
433 }
434
435 /*
436 * The following routine will lock n inodes in exclusive mode. We assume the
437 * caller calls us with the inodes in i_ino order.
438 *
439 * We need to detect deadlock where an inode that we lock is in the AIL and we
440 * start waiting for another inode that is locked by a thread in a long running
441 * transaction (such as truncate). This can result in deadlock since the long
442 * running trans might need to wait for the inode we just locked in order to
443 * push the tail and free space in the log.
444 *
445 * xfs_lock_inodes() can only be used to lock one type of lock at a time -
446 * the iolock, the mmaplock or the ilock, but not more than one at a time. If we
447 * lock more than one at a time, lockdep will report false positives saying we
448 * have violated locking orders.
449 */
450 static void
xfs_lock_inodes(struct xfs_inode ** ips,int inodes,uint lock_mode)451 xfs_lock_inodes(
452 struct xfs_inode **ips,
453 int inodes,
454 uint lock_mode)
455 {
456 int attempts = 0;
457 uint i;
458 int j;
459 bool try_lock;
460 struct xfs_log_item *lp;
461
462 /*
463 * Currently supports between 2 and 5 inodes with exclusive locking. We
464 * support an arbitrary depth of locking here, but absolute limits on
465 * inodes depend on the type of locking and the limits placed by
466 * lockdep annotations in xfs_lock_inumorder. These are all checked by
467 * the asserts.
468 */
469 ASSERT(ips && inodes >= 2 && inodes <= 5);
470 ASSERT(lock_mode & (XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL |
471 XFS_ILOCK_EXCL));
472 ASSERT(!(lock_mode & (XFS_IOLOCK_SHARED | XFS_MMAPLOCK_SHARED |
473 XFS_ILOCK_SHARED)));
474 ASSERT(!(lock_mode & XFS_MMAPLOCK_EXCL) ||
475 inodes <= XFS_MMAPLOCK_MAX_SUBCLASS + 1);
476 ASSERT(!(lock_mode & XFS_ILOCK_EXCL) ||
477 inodes <= XFS_ILOCK_MAX_SUBCLASS + 1);
478
479 if (lock_mode & XFS_IOLOCK_EXCL) {
480 ASSERT(!(lock_mode & (XFS_MMAPLOCK_EXCL | XFS_ILOCK_EXCL)));
481 } else if (lock_mode & XFS_MMAPLOCK_EXCL)
482 ASSERT(!(lock_mode & XFS_ILOCK_EXCL));
483
484 again:
485 try_lock = false;
486 i = 0;
487 for (; i < inodes; i++) {
488 ASSERT(ips[i]);
489
490 if (i && (ips[i] == ips[i - 1])) /* Already locked */
491 continue;
492
493 /*
494 * If try_lock is not set yet, make sure all locked inodes are
495 * not in the AIL. If any are, set try_lock to be used later.
496 */
497 if (!try_lock) {
498 for (j = (i - 1); j >= 0 && !try_lock; j--) {
499 lp = &ips[j]->i_itemp->ili_item;
500 if (lp && test_bit(XFS_LI_IN_AIL, &lp->li_flags))
501 try_lock = true;
502 }
503 }
504
505 /*
506 * If any of the previous locks we have locked is in the AIL,
507 * we must TRY to get the second and subsequent locks. If
508 * we can't get any, we must release all we have
509 * and try again.
510 */
511 if (!try_lock) {
512 xfs_ilock(ips[i], xfs_lock_inumorder(lock_mode, i));
513 continue;
514 }
515
516 /* try_lock means we have an inode locked that is in the AIL. */
517 ASSERT(i != 0);
518 if (xfs_ilock_nowait(ips[i], xfs_lock_inumorder(lock_mode, i)))
519 continue;
520
521 /*
522 * Unlock all previous guys and try again. xfs_iunlock will try
523 * to push the tail if the inode is in the AIL.
524 */
525 attempts++;
526 for (j = i - 1; j >= 0; j--) {
527 /*
528 * Check to see if we've already unlocked this one. Not
529 * the first one going back, and the inode ptr is the
530 * same.
531 */
532 if (j != (i - 1) && ips[j] == ips[j + 1])
533 continue;
534
535 xfs_iunlock(ips[j], lock_mode);
536 }
537
538 if ((attempts % 5) == 0) {
539 delay(1); /* Don't just spin the CPU */
540 }
541 goto again;
542 }
543 }
544
545 /*
546 * xfs_lock_two_inodes() can only be used to lock ilock. The iolock and
547 * mmaplock must be double-locked separately since we use i_rwsem and
548 * invalidate_lock for that. We now support taking one lock EXCL and the
549 * other SHARED.
550 */
551 void
xfs_lock_two_inodes(struct xfs_inode * ip0,uint ip0_mode,struct xfs_inode * ip1,uint ip1_mode)552 xfs_lock_two_inodes(
553 struct xfs_inode *ip0,
554 uint ip0_mode,
555 struct xfs_inode *ip1,
556 uint ip1_mode)
557 {
558 int attempts = 0;
559 struct xfs_log_item *lp;
560
561 ASSERT(hweight32(ip0_mode) == 1);
562 ASSERT(hweight32(ip1_mode) == 1);
563 ASSERT(!(ip0_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
564 ASSERT(!(ip1_mode & (XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL)));
565 ASSERT(!(ip0_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)));
566 ASSERT(!(ip1_mode & (XFS_MMAPLOCK_SHARED|XFS_MMAPLOCK_EXCL)));
567 ASSERT(ip0->i_ino != ip1->i_ino);
568
569 if (ip0->i_ino > ip1->i_ino) {
570 swap(ip0, ip1);
571 swap(ip0_mode, ip1_mode);
572 }
573
574 again:
575 xfs_ilock(ip0, xfs_lock_inumorder(ip0_mode, 0));
576
577 /*
578 * If the first lock we have locked is in the AIL, we must TRY to get
579 * the second lock. If we can't get it, we must release the first one
580 * and try again.
581 */
582 lp = &ip0->i_itemp->ili_item;
583 if (lp && test_bit(XFS_LI_IN_AIL, &lp->li_flags)) {
584 if (!xfs_ilock_nowait(ip1, xfs_lock_inumorder(ip1_mode, 1))) {
585 xfs_iunlock(ip0, ip0_mode);
586 if ((++attempts % 5) == 0)
587 delay(1); /* Don't just spin the CPU */
588 goto again;
589 }
590 } else {
591 xfs_ilock(ip1, xfs_lock_inumorder(ip1_mode, 1));
592 }
593 }
594
595 uint
xfs_ip2xflags(struct xfs_inode * ip)596 xfs_ip2xflags(
597 struct xfs_inode *ip)
598 {
599 uint flags = 0;
600
601 if (ip->i_diflags & XFS_DIFLAG_ANY) {
602 if (ip->i_diflags & XFS_DIFLAG_REALTIME)
603 flags |= FS_XFLAG_REALTIME;
604 if (ip->i_diflags & XFS_DIFLAG_PREALLOC)
605 flags |= FS_XFLAG_PREALLOC;
606 if (ip->i_diflags & XFS_DIFLAG_IMMUTABLE)
607 flags |= FS_XFLAG_IMMUTABLE;
608 if (ip->i_diflags & XFS_DIFLAG_APPEND)
609 flags |= FS_XFLAG_APPEND;
610 if (ip->i_diflags & XFS_DIFLAG_SYNC)
611 flags |= FS_XFLAG_SYNC;
612 if (ip->i_diflags & XFS_DIFLAG_NOATIME)
613 flags |= FS_XFLAG_NOATIME;
614 if (ip->i_diflags & XFS_DIFLAG_NODUMP)
615 flags |= FS_XFLAG_NODUMP;
616 if (ip->i_diflags & XFS_DIFLAG_RTINHERIT)
617 flags |= FS_XFLAG_RTINHERIT;
618 if (ip->i_diflags & XFS_DIFLAG_PROJINHERIT)
619 flags |= FS_XFLAG_PROJINHERIT;
620 if (ip->i_diflags & XFS_DIFLAG_NOSYMLINKS)
621 flags |= FS_XFLAG_NOSYMLINKS;
622 if (ip->i_diflags & XFS_DIFLAG_EXTSIZE)
623 flags |= FS_XFLAG_EXTSIZE;
624 if (ip->i_diflags & XFS_DIFLAG_EXTSZINHERIT)
625 flags |= FS_XFLAG_EXTSZINHERIT;
626 if (ip->i_diflags & XFS_DIFLAG_NODEFRAG)
627 flags |= FS_XFLAG_NODEFRAG;
628 if (ip->i_diflags & XFS_DIFLAG_FILESTREAM)
629 flags |= FS_XFLAG_FILESTREAM;
630 }
631
632 if (ip->i_diflags2 & XFS_DIFLAG2_ANY) {
633 if (ip->i_diflags2 & XFS_DIFLAG2_DAX)
634 flags |= FS_XFLAG_DAX;
635 if (ip->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE)
636 flags |= FS_XFLAG_COWEXTSIZE;
637 }
638
639 if (xfs_inode_has_attr_fork(ip))
640 flags |= FS_XFLAG_HASATTR;
641 return flags;
642 }
643
644 /*
645 * Lookups up an inode from "name". If ci_name is not NULL, then a CI match
646 * is allowed, otherwise it has to be an exact match. If a CI match is found,
647 * ci_name->name will point to a the actual name (caller must free) or
648 * will be set to NULL if an exact match is found.
649 */
650 int
xfs_lookup(struct xfs_inode * dp,const struct xfs_name * name,struct xfs_inode ** ipp,struct xfs_name * ci_name)651 xfs_lookup(
652 struct xfs_inode *dp,
653 const struct xfs_name *name,
654 struct xfs_inode **ipp,
655 struct xfs_name *ci_name)
656 {
657 xfs_ino_t inum;
658 int error;
659
660 trace_xfs_lookup(dp, name);
661
662 if (xfs_is_shutdown(dp->i_mount))
663 return -EIO;
664
665 error = xfs_dir_lookup(NULL, dp, name, &inum, ci_name);
666 if (error)
667 goto out_unlock;
668
669 error = xfs_iget(dp->i_mount, NULL, inum, 0, 0, ipp);
670 if (error)
671 goto out_free_name;
672
673 return 0;
674
675 out_free_name:
676 if (ci_name)
677 kmem_free(ci_name->name);
678 out_unlock:
679 *ipp = NULL;
680 return error;
681 }
682
683 /* Propagate di_flags from a parent inode to a child inode. */
684 static void
xfs_inode_inherit_flags(struct xfs_inode * ip,const struct xfs_inode * pip)685 xfs_inode_inherit_flags(
686 struct xfs_inode *ip,
687 const struct xfs_inode *pip)
688 {
689 unsigned int di_flags = 0;
690 xfs_failaddr_t failaddr;
691 umode_t mode = VFS_I(ip)->i_mode;
692
693 if (S_ISDIR(mode)) {
694 if (pip->i_diflags & XFS_DIFLAG_RTINHERIT)
695 di_flags |= XFS_DIFLAG_RTINHERIT;
696 if (pip->i_diflags & XFS_DIFLAG_EXTSZINHERIT) {
697 di_flags |= XFS_DIFLAG_EXTSZINHERIT;
698 ip->i_extsize = pip->i_extsize;
699 }
700 if (pip->i_diflags & XFS_DIFLAG_PROJINHERIT)
701 di_flags |= XFS_DIFLAG_PROJINHERIT;
702 } else if (S_ISREG(mode)) {
703 if ((pip->i_diflags & XFS_DIFLAG_RTINHERIT) &&
704 xfs_has_realtime(ip->i_mount))
705 di_flags |= XFS_DIFLAG_REALTIME;
706 if (pip->i_diflags & XFS_DIFLAG_EXTSZINHERIT) {
707 di_flags |= XFS_DIFLAG_EXTSIZE;
708 ip->i_extsize = pip->i_extsize;
709 }
710 }
711 if ((pip->i_diflags & XFS_DIFLAG_NOATIME) &&
712 xfs_inherit_noatime)
713 di_flags |= XFS_DIFLAG_NOATIME;
714 if ((pip->i_diflags & XFS_DIFLAG_NODUMP) &&
715 xfs_inherit_nodump)
716 di_flags |= XFS_DIFLAG_NODUMP;
717 if ((pip->i_diflags & XFS_DIFLAG_SYNC) &&
718 xfs_inherit_sync)
719 di_flags |= XFS_DIFLAG_SYNC;
720 if ((pip->i_diflags & XFS_DIFLAG_NOSYMLINKS) &&
721 xfs_inherit_nosymlinks)
722 di_flags |= XFS_DIFLAG_NOSYMLINKS;
723 if ((pip->i_diflags & XFS_DIFLAG_NODEFRAG) &&
724 xfs_inherit_nodefrag)
725 di_flags |= XFS_DIFLAG_NODEFRAG;
726 if (pip->i_diflags & XFS_DIFLAG_FILESTREAM)
727 di_flags |= XFS_DIFLAG_FILESTREAM;
728
729 ip->i_diflags |= di_flags;
730
731 /*
732 * Inode verifiers on older kernels only check that the extent size
733 * hint is an integer multiple of the rt extent size on realtime files.
734 * They did not check the hint alignment on a directory with both
735 * rtinherit and extszinherit flags set. If the misaligned hint is
736 * propagated from a directory into a new realtime file, new file
737 * allocations will fail due to math errors in the rt allocator and/or
738 * trip the verifiers. Validate the hint settings in the new file so
739 * that we don't let broken hints propagate.
740 */
741 failaddr = xfs_inode_validate_extsize(ip->i_mount, ip->i_extsize,
742 VFS_I(ip)->i_mode, ip->i_diflags);
743 if (failaddr) {
744 ip->i_diflags &= ~(XFS_DIFLAG_EXTSIZE |
745 XFS_DIFLAG_EXTSZINHERIT);
746 ip->i_extsize = 0;
747 }
748 }
749
750 /* Propagate di_flags2 from a parent inode to a child inode. */
751 static void
xfs_inode_inherit_flags2(struct xfs_inode * ip,const struct xfs_inode * pip)752 xfs_inode_inherit_flags2(
753 struct xfs_inode *ip,
754 const struct xfs_inode *pip)
755 {
756 xfs_failaddr_t failaddr;
757
758 if (pip->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE) {
759 ip->i_diflags2 |= XFS_DIFLAG2_COWEXTSIZE;
760 ip->i_cowextsize = pip->i_cowextsize;
761 }
762 if (pip->i_diflags2 & XFS_DIFLAG2_DAX)
763 ip->i_diflags2 |= XFS_DIFLAG2_DAX;
764
765 /* Don't let invalid cowextsize hints propagate. */
766 failaddr = xfs_inode_validate_cowextsize(ip->i_mount, ip->i_cowextsize,
767 VFS_I(ip)->i_mode, ip->i_diflags, ip->i_diflags2);
768 if (failaddr) {
769 ip->i_diflags2 &= ~XFS_DIFLAG2_COWEXTSIZE;
770 ip->i_cowextsize = 0;
771 }
772 }
773
774 /*
775 * Initialise a newly allocated inode and return the in-core inode to the
776 * caller locked exclusively.
777 */
778 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)779 xfs_init_new_inode(
780 struct user_namespace *mnt_userns,
781 struct xfs_trans *tp,
782 struct xfs_inode *pip,
783 xfs_ino_t ino,
784 umode_t mode,
785 xfs_nlink_t nlink,
786 dev_t rdev,
787 prid_t prid,
788 bool init_xattrs,
789 struct xfs_inode **ipp)
790 {
791 struct inode *dir = pip ? VFS_I(pip) : NULL;
792 struct xfs_mount *mp = tp->t_mountp;
793 struct xfs_inode *ip;
794 unsigned int flags;
795 int error;
796 struct timespec64 tv;
797 struct inode *inode;
798
799 /*
800 * Protect against obviously corrupt allocation btree records. Later
801 * xfs_iget checks will catch re-allocation of other active in-memory
802 * and on-disk inodes. If we don't catch reallocating the parent inode
803 * here we will deadlock in xfs_iget() so we have to do these checks
804 * first.
805 */
806 if ((pip && ino == pip->i_ino) || !xfs_verify_dir_ino(mp, ino)) {
807 xfs_alert(mp, "Allocated a known in-use inode 0x%llx!", ino);
808 return -EFSCORRUPTED;
809 }
810
811 /*
812 * Get the in-core inode with the lock held exclusively to prevent
813 * others from looking at until we're done.
814 */
815 error = xfs_iget(mp, tp, ino, XFS_IGET_CREATE, XFS_ILOCK_EXCL, &ip);
816 if (error)
817 return error;
818
819 ASSERT(ip != NULL);
820 inode = VFS_I(ip);
821 set_nlink(inode, nlink);
822 inode->i_rdev = rdev;
823 ip->i_projid = prid;
824
825 if (dir && !(dir->i_mode & S_ISGID) && xfs_has_grpid(mp)) {
826 inode_fsuid_set(inode, mnt_userns);
827 inode->i_gid = dir->i_gid;
828 inode->i_mode = mode;
829 } else {
830 inode_init_owner(mnt_userns, inode, dir, mode);
831 }
832
833 /*
834 * If the group ID of the new file does not match the effective group
835 * ID or one of the supplementary group IDs, the S_ISGID bit is cleared
836 * (and only if the irix_sgid_inherit compatibility variable is set).
837 */
838 if (irix_sgid_inherit && (inode->i_mode & S_ISGID) &&
839 !vfsgid_in_group_p(i_gid_into_vfsgid(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 xfs_ifork_init_attr(ip, 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_inode_has_attr_fork(ip)) {
1766 error = xfs_attr_inactive(ip);
1767 if (error)
1768 goto out;
1769 }
1770
1771 ASSERT(ip->i_forkoff == 0);
1772
1773 /*
1774 * Free the inode.
1775 */
1776 xfs_inactive_ifree(ip);
1777
1778 out:
1779 /*
1780 * We're done making metadata updates for this inode, so we can release
1781 * the attached dquots.
1782 */
1783 xfs_qm_dqdetach(ip);
1784 }
1785
1786 /*
1787 * In-Core Unlinked List Lookups
1788 * =============================
1789 *
1790 * Every inode is supposed to be reachable from some other piece of metadata
1791 * with the exception of the root directory. Inodes with a connection to a
1792 * file descriptor but not linked from anywhere in the on-disk directory tree
1793 * are collectively known as unlinked inodes, though the filesystem itself
1794 * maintains links to these inodes so that on-disk metadata are consistent.
1795 *
1796 * XFS implements a per-AG on-disk hash table of unlinked inodes. The AGI
1797 * header contains a number of buckets that point to an inode, and each inode
1798 * record has a pointer to the next inode in the hash chain. This
1799 * singly-linked list causes scaling problems in the iunlink remove function
1800 * because we must walk that list to find the inode that points to the inode
1801 * being removed from the unlinked hash bucket list.
1802 *
1803 * Hence we keep an in-memory double linked list to link each inode on an
1804 * unlinked list. Because there are 64 unlinked lists per AGI, keeping pointer
1805 * based lists would require having 64 list heads in the perag, one for each
1806 * list. This is expensive in terms of memory (think millions of AGs) and cache
1807 * misses on lookups. Instead, use the fact that inodes on the unlinked list
1808 * must be referenced at the VFS level to keep them on the list and hence we
1809 * have an existence guarantee for inodes on the unlinked list.
1810 *
1811 * Given we have an existence guarantee, we can use lockless inode cache lookups
1812 * to resolve aginos to xfs inodes. This means we only need 8 bytes per inode
1813 * for the double linked unlinked list, and we don't need any extra locking to
1814 * keep the list safe as all manipulations are done under the AGI buffer lock.
1815 * Keeping the list up to date does not require memory allocation, just finding
1816 * the XFS inode and updating the next/prev unlinked list aginos.
1817 */
1818
1819 /*
1820 * Find an inode on the unlinked list. This does not take references to the
1821 * inode as we have existence guarantees by holding the AGI buffer lock and that
1822 * only unlinked, referenced inodes can be on the unlinked inode list. If we
1823 * don't find the inode in cache, then let the caller handle the situation.
1824 */
1825 static struct xfs_inode *
xfs_iunlink_lookup(struct xfs_perag * pag,xfs_agino_t agino)1826 xfs_iunlink_lookup(
1827 struct xfs_perag *pag,
1828 xfs_agino_t agino)
1829 {
1830 struct xfs_inode *ip;
1831
1832 rcu_read_lock();
1833 ip = radix_tree_lookup(&pag->pag_ici_root, agino);
1834
1835 /*
1836 * Inode not in memory or in RCU freeing limbo should not happen.
1837 * Warn about this and let the caller handle the failure.
1838 */
1839 if (WARN_ON_ONCE(!ip || !ip->i_ino)) {
1840 rcu_read_unlock();
1841 return NULL;
1842 }
1843 ASSERT(!xfs_iflags_test(ip, XFS_IRECLAIMABLE | XFS_IRECLAIM));
1844 rcu_read_unlock();
1845 return ip;
1846 }
1847
1848 /* Update the prev pointer of the next agino. */
1849 static int
xfs_iunlink_update_backref(struct xfs_perag * pag,xfs_agino_t prev_agino,xfs_agino_t next_agino)1850 xfs_iunlink_update_backref(
1851 struct xfs_perag *pag,
1852 xfs_agino_t prev_agino,
1853 xfs_agino_t next_agino)
1854 {
1855 struct xfs_inode *ip;
1856
1857 /* No update necessary if we are at the end of the list. */
1858 if (next_agino == NULLAGINO)
1859 return 0;
1860
1861 ip = xfs_iunlink_lookup(pag, next_agino);
1862 if (!ip)
1863 return -EFSCORRUPTED;
1864 ip->i_prev_unlinked = prev_agino;
1865 return 0;
1866 }
1867
1868 /*
1869 * Point the AGI unlinked bucket at an inode and log the results. The caller
1870 * is responsible for validating the old value.
1871 */
1872 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)1873 xfs_iunlink_update_bucket(
1874 struct xfs_trans *tp,
1875 struct xfs_perag *pag,
1876 struct xfs_buf *agibp,
1877 unsigned int bucket_index,
1878 xfs_agino_t new_agino)
1879 {
1880 struct xfs_agi *agi = agibp->b_addr;
1881 xfs_agino_t old_value;
1882 int offset;
1883
1884 ASSERT(xfs_verify_agino_or_null(pag, new_agino));
1885
1886 old_value = be32_to_cpu(agi->agi_unlinked[bucket_index]);
1887 trace_xfs_iunlink_update_bucket(tp->t_mountp, pag->pag_agno, bucket_index,
1888 old_value, new_agino);
1889
1890 /*
1891 * We should never find the head of the list already set to the value
1892 * passed in because either we're adding or removing ourselves from the
1893 * head of the list.
1894 */
1895 if (old_value == new_agino) {
1896 xfs_buf_mark_corrupt(agibp);
1897 return -EFSCORRUPTED;
1898 }
1899
1900 agi->agi_unlinked[bucket_index] = cpu_to_be32(new_agino);
1901 offset = offsetof(struct xfs_agi, agi_unlinked) +
1902 (sizeof(xfs_agino_t) * bucket_index);
1903 xfs_trans_log_buf(tp, agibp, offset, offset + sizeof(xfs_agino_t) - 1);
1904 return 0;
1905 }
1906
1907 static int
xfs_iunlink_insert_inode(struct xfs_trans * tp,struct xfs_perag * pag,struct xfs_buf * agibp,struct xfs_inode * ip)1908 xfs_iunlink_insert_inode(
1909 struct xfs_trans *tp,
1910 struct xfs_perag *pag,
1911 struct xfs_buf *agibp,
1912 struct xfs_inode *ip)
1913 {
1914 struct xfs_mount *mp = tp->t_mountp;
1915 struct xfs_agi *agi = agibp->b_addr;
1916 xfs_agino_t next_agino;
1917 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
1918 short bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
1919 int error;
1920
1921 /*
1922 * Get the index into the agi hash table for the list this inode will
1923 * go on. Make sure the pointer isn't garbage and that this inode
1924 * isn't already on the list.
1925 */
1926 next_agino = be32_to_cpu(agi->agi_unlinked[bucket_index]);
1927 if (next_agino == agino ||
1928 !xfs_verify_agino_or_null(pag, next_agino)) {
1929 xfs_buf_mark_corrupt(agibp);
1930 return -EFSCORRUPTED;
1931 }
1932
1933 /*
1934 * Update the prev pointer in the next inode to point back to this
1935 * inode.
1936 */
1937 error = xfs_iunlink_update_backref(pag, agino, next_agino);
1938 if (error)
1939 return error;
1940
1941 if (next_agino != NULLAGINO) {
1942 /*
1943 * There is already another inode in the bucket, so point this
1944 * inode to the current head of the list.
1945 */
1946 error = xfs_iunlink_log_inode(tp, ip, pag, next_agino);
1947 if (error)
1948 return error;
1949 ip->i_next_unlinked = next_agino;
1950 }
1951
1952 /* Point the head of the list to point to this inode. */
1953 return xfs_iunlink_update_bucket(tp, pag, agibp, bucket_index, agino);
1954 }
1955
1956 /*
1957 * This is called when the inode's link count has gone to 0 or we are creating
1958 * a tmpfile via O_TMPFILE. The inode @ip must have nlink == 0.
1959 *
1960 * We place the on-disk inode on a list in the AGI. It will be pulled from this
1961 * list when the inode is freed.
1962 */
1963 STATIC int
xfs_iunlink(struct xfs_trans * tp,struct xfs_inode * ip)1964 xfs_iunlink(
1965 struct xfs_trans *tp,
1966 struct xfs_inode *ip)
1967 {
1968 struct xfs_mount *mp = tp->t_mountp;
1969 struct xfs_perag *pag;
1970 struct xfs_buf *agibp;
1971 int error;
1972
1973 ASSERT(VFS_I(ip)->i_nlink == 0);
1974 ASSERT(VFS_I(ip)->i_mode != 0);
1975 trace_xfs_iunlink(ip);
1976
1977 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
1978
1979 /* Get the agi buffer first. It ensures lock ordering on the list. */
1980 error = xfs_read_agi(pag, tp, &agibp);
1981 if (error)
1982 goto out;
1983
1984 error = xfs_iunlink_insert_inode(tp, pag, agibp, ip);
1985 out:
1986 xfs_perag_put(pag);
1987 return error;
1988 }
1989
1990 static int
xfs_iunlink_remove_inode(struct xfs_trans * tp,struct xfs_perag * pag,struct xfs_buf * agibp,struct xfs_inode * ip)1991 xfs_iunlink_remove_inode(
1992 struct xfs_trans *tp,
1993 struct xfs_perag *pag,
1994 struct xfs_buf *agibp,
1995 struct xfs_inode *ip)
1996 {
1997 struct xfs_mount *mp = tp->t_mountp;
1998 struct xfs_agi *agi = agibp->b_addr;
1999 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ip->i_ino);
2000 xfs_agino_t head_agino;
2001 short bucket_index = agino % XFS_AGI_UNLINKED_BUCKETS;
2002 int error;
2003
2004 trace_xfs_iunlink_remove(ip);
2005
2006 /*
2007 * Get the index into the agi hash table for the list this inode will
2008 * go on. Make sure the head pointer isn't garbage.
2009 */
2010 head_agino = be32_to_cpu(agi->agi_unlinked[bucket_index]);
2011 if (!xfs_verify_agino(pag, head_agino)) {
2012 XFS_CORRUPTION_ERROR(__func__, XFS_ERRLEVEL_LOW, mp,
2013 agi, sizeof(*agi));
2014 return -EFSCORRUPTED;
2015 }
2016
2017 /*
2018 * Set our inode's next_unlinked pointer to NULL and then return
2019 * the old pointer value so that we can update whatever was previous
2020 * to us in the list to point to whatever was next in the list.
2021 */
2022 error = xfs_iunlink_log_inode(tp, ip, pag, NULLAGINO);
2023 if (error)
2024 return error;
2025
2026 /*
2027 * Update the prev pointer in the next inode to point back to previous
2028 * inode in the chain.
2029 */
2030 error = xfs_iunlink_update_backref(pag, ip->i_prev_unlinked,
2031 ip->i_next_unlinked);
2032 if (error)
2033 return error;
2034
2035 if (head_agino != agino) {
2036 struct xfs_inode *prev_ip;
2037
2038 prev_ip = xfs_iunlink_lookup(pag, ip->i_prev_unlinked);
2039 if (!prev_ip)
2040 return -EFSCORRUPTED;
2041
2042 error = xfs_iunlink_log_inode(tp, prev_ip, pag,
2043 ip->i_next_unlinked);
2044 prev_ip->i_next_unlinked = ip->i_next_unlinked;
2045 } else {
2046 /* Point the head of the list to the next unlinked inode. */
2047 error = xfs_iunlink_update_bucket(tp, pag, agibp, bucket_index,
2048 ip->i_next_unlinked);
2049 }
2050
2051 ip->i_next_unlinked = NULLAGINO;
2052 ip->i_prev_unlinked = NULLAGINO;
2053 return error;
2054 }
2055
2056 /*
2057 * Pull the on-disk inode from the AGI unlinked list.
2058 */
2059 STATIC int
xfs_iunlink_remove(struct xfs_trans * tp,struct xfs_perag * pag,struct xfs_inode * ip)2060 xfs_iunlink_remove(
2061 struct xfs_trans *tp,
2062 struct xfs_perag *pag,
2063 struct xfs_inode *ip)
2064 {
2065 struct xfs_buf *agibp;
2066 int error;
2067
2068 trace_xfs_iunlink_remove(ip);
2069
2070 /* Get the agi buffer first. It ensures lock ordering on the list. */
2071 error = xfs_read_agi(pag, tp, &agibp);
2072 if (error)
2073 return error;
2074
2075 return xfs_iunlink_remove_inode(tp, pag, agibp, ip);
2076 }
2077
2078 /*
2079 * Look up the inode number specified and if it is not already marked XFS_ISTALE
2080 * mark it stale. We should only find clean inodes in this lookup that aren't
2081 * already stale.
2082 */
2083 static void
xfs_ifree_mark_inode_stale(struct xfs_perag * pag,struct xfs_inode * free_ip,xfs_ino_t inum)2084 xfs_ifree_mark_inode_stale(
2085 struct xfs_perag *pag,
2086 struct xfs_inode *free_ip,
2087 xfs_ino_t inum)
2088 {
2089 struct xfs_mount *mp = pag->pag_mount;
2090 struct xfs_inode_log_item *iip;
2091 struct xfs_inode *ip;
2092
2093 retry:
2094 rcu_read_lock();
2095 ip = radix_tree_lookup(&pag->pag_ici_root, XFS_INO_TO_AGINO(mp, inum));
2096
2097 /* Inode not in memory, nothing to do */
2098 if (!ip) {
2099 rcu_read_unlock();
2100 return;
2101 }
2102
2103 /*
2104 * because this is an RCU protected lookup, we could find a recently
2105 * freed or even reallocated inode during the lookup. We need to check
2106 * under the i_flags_lock for a valid inode here. Skip it if it is not
2107 * valid, the wrong inode or stale.
2108 */
2109 spin_lock(&ip->i_flags_lock);
2110 if (ip->i_ino != inum || __xfs_iflags_test(ip, XFS_ISTALE))
2111 goto out_iflags_unlock;
2112
2113 /*
2114 * Don't try to lock/unlock the current inode, but we _cannot_ skip the
2115 * other inodes that we did not find in the list attached to the buffer
2116 * and are not already marked stale. If we can't lock it, back off and
2117 * retry.
2118 */
2119 if (ip != free_ip) {
2120 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL)) {
2121 spin_unlock(&ip->i_flags_lock);
2122 rcu_read_unlock();
2123 delay(1);
2124 goto retry;
2125 }
2126 }
2127 ip->i_flags |= XFS_ISTALE;
2128
2129 /*
2130 * If the inode is flushing, it is already attached to the buffer. All
2131 * we needed to do here is mark the inode stale so buffer IO completion
2132 * will remove it from the AIL.
2133 */
2134 iip = ip->i_itemp;
2135 if (__xfs_iflags_test(ip, XFS_IFLUSHING)) {
2136 ASSERT(!list_empty(&iip->ili_item.li_bio_list));
2137 ASSERT(iip->ili_last_fields);
2138 goto out_iunlock;
2139 }
2140
2141 /*
2142 * Inodes not attached to the buffer can be released immediately.
2143 * Everything else has to go through xfs_iflush_abort() on journal
2144 * commit as the flock synchronises removal of the inode from the
2145 * cluster buffer against inode reclaim.
2146 */
2147 if (!iip || list_empty(&iip->ili_item.li_bio_list))
2148 goto out_iunlock;
2149
2150 __xfs_iflags_set(ip, XFS_IFLUSHING);
2151 spin_unlock(&ip->i_flags_lock);
2152 rcu_read_unlock();
2153
2154 /* we have a dirty inode in memory that has not yet been flushed. */
2155 spin_lock(&iip->ili_lock);
2156 iip->ili_last_fields = iip->ili_fields;
2157 iip->ili_fields = 0;
2158 iip->ili_fsync_fields = 0;
2159 spin_unlock(&iip->ili_lock);
2160 ASSERT(iip->ili_last_fields);
2161
2162 if (ip != free_ip)
2163 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2164 return;
2165
2166 out_iunlock:
2167 if (ip != free_ip)
2168 xfs_iunlock(ip, XFS_ILOCK_EXCL);
2169 out_iflags_unlock:
2170 spin_unlock(&ip->i_flags_lock);
2171 rcu_read_unlock();
2172 }
2173
2174 /*
2175 * A big issue when freeing the inode cluster is that we _cannot_ skip any
2176 * inodes that are in memory - they all must be marked stale and attached to
2177 * the cluster buffer.
2178 */
2179 static int
xfs_ifree_cluster(struct xfs_trans * tp,struct xfs_perag * pag,struct xfs_inode * free_ip,struct xfs_icluster * xic)2180 xfs_ifree_cluster(
2181 struct xfs_trans *tp,
2182 struct xfs_perag *pag,
2183 struct xfs_inode *free_ip,
2184 struct xfs_icluster *xic)
2185 {
2186 struct xfs_mount *mp = free_ip->i_mount;
2187 struct xfs_ino_geometry *igeo = M_IGEO(mp);
2188 struct xfs_buf *bp;
2189 xfs_daddr_t blkno;
2190 xfs_ino_t inum = xic->first_ino;
2191 int nbufs;
2192 int i, j;
2193 int ioffset;
2194 int error;
2195
2196 nbufs = igeo->ialloc_blks / igeo->blocks_per_cluster;
2197
2198 for (j = 0; j < nbufs; j++, inum += igeo->inodes_per_cluster) {
2199 /*
2200 * The allocation bitmap tells us which inodes of the chunk were
2201 * physically allocated. Skip the cluster if an inode falls into
2202 * a sparse region.
2203 */
2204 ioffset = inum - xic->first_ino;
2205 if ((xic->alloc & XFS_INOBT_MASK(ioffset)) == 0) {
2206 ASSERT(ioffset % igeo->inodes_per_cluster == 0);
2207 continue;
2208 }
2209
2210 blkno = XFS_AGB_TO_DADDR(mp, XFS_INO_TO_AGNO(mp, inum),
2211 XFS_INO_TO_AGBNO(mp, inum));
2212
2213 /*
2214 * We obtain and lock the backing buffer first in the process
2215 * here to ensure dirty inodes attached to the buffer remain in
2216 * the flushing state while we mark them stale.
2217 *
2218 * If we scan the in-memory inodes first, then buffer IO can
2219 * complete before we get a lock on it, and hence we may fail
2220 * to mark all the active inodes on the buffer stale.
2221 */
2222 error = xfs_trans_get_buf(tp, mp->m_ddev_targp, blkno,
2223 mp->m_bsize * igeo->blocks_per_cluster,
2224 XBF_UNMAPPED, &bp);
2225 if (error)
2226 return error;
2227
2228 /*
2229 * This buffer may not have been correctly initialised as we
2230 * didn't read it from disk. That's not important because we are
2231 * only using to mark the buffer as stale in the log, and to
2232 * attach stale cached inodes on it. That means it will never be
2233 * dispatched for IO. If it is, we want to know about it, and we
2234 * want it to fail. We can acheive this by adding a write
2235 * verifier to the buffer.
2236 */
2237 bp->b_ops = &xfs_inode_buf_ops;
2238
2239 /*
2240 * Now we need to set all the cached clean inodes as XFS_ISTALE,
2241 * too. This requires lookups, and will skip inodes that we've
2242 * already marked XFS_ISTALE.
2243 */
2244 for (i = 0; i < igeo->inodes_per_cluster; i++)
2245 xfs_ifree_mark_inode_stale(pag, free_ip, inum + i);
2246
2247 xfs_trans_stale_inode_buf(tp, bp);
2248 xfs_trans_binval(tp, bp);
2249 }
2250 return 0;
2251 }
2252
2253 /*
2254 * This is called to return an inode to the inode free list. The inode should
2255 * already be truncated to 0 length and have no pages associated with it. This
2256 * routine also assumes that the inode is already a part of the transaction.
2257 *
2258 * The on-disk copy of the inode will have been added to the list of unlinked
2259 * inodes in the AGI. We need to remove the inode from that list atomically with
2260 * respect to freeing it here.
2261 */
2262 int
xfs_ifree(struct xfs_trans * tp,struct xfs_inode * ip)2263 xfs_ifree(
2264 struct xfs_trans *tp,
2265 struct xfs_inode *ip)
2266 {
2267 struct xfs_mount *mp = ip->i_mount;
2268 struct xfs_perag *pag;
2269 struct xfs_icluster xic = { 0 };
2270 struct xfs_inode_log_item *iip = ip->i_itemp;
2271 int error;
2272
2273 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
2274 ASSERT(VFS_I(ip)->i_nlink == 0);
2275 ASSERT(ip->i_df.if_nextents == 0);
2276 ASSERT(ip->i_disk_size == 0 || !S_ISREG(VFS_I(ip)->i_mode));
2277 ASSERT(ip->i_nblocks == 0);
2278
2279 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ip->i_ino));
2280
2281 /*
2282 * Free the inode first so that we guarantee that the AGI lock is going
2283 * to be taken before we remove the inode from the unlinked list. This
2284 * makes the AGI lock -> unlinked list modification order the same as
2285 * used in O_TMPFILE creation.
2286 */
2287 error = xfs_difree(tp, pag, ip->i_ino, &xic);
2288 if (error)
2289 goto out;
2290
2291 error = xfs_iunlink_remove(tp, pag, ip);
2292 if (error)
2293 goto out;
2294
2295 /*
2296 * Free any local-format data sitting around before we reset the
2297 * data fork to extents format. Note that the attr fork data has
2298 * already been freed by xfs_attr_inactive.
2299 */
2300 if (ip->i_df.if_format == XFS_DINODE_FMT_LOCAL) {
2301 kmem_free(ip->i_df.if_u1.if_data);
2302 ip->i_df.if_u1.if_data = NULL;
2303 ip->i_df.if_bytes = 0;
2304 }
2305
2306 VFS_I(ip)->i_mode = 0; /* mark incore inode as free */
2307 ip->i_diflags = 0;
2308 ip->i_diflags2 = mp->m_ino_geo.new_diflags2;
2309 ip->i_forkoff = 0; /* mark the attr fork not in use */
2310 ip->i_df.if_format = XFS_DINODE_FMT_EXTENTS;
2311 if (xfs_iflags_test(ip, XFS_IPRESERVE_DM_FIELDS))
2312 xfs_iflags_clear(ip, XFS_IPRESERVE_DM_FIELDS);
2313
2314 /* Don't attempt to replay owner changes for a deleted inode */
2315 spin_lock(&iip->ili_lock);
2316 iip->ili_fields &= ~(XFS_ILOG_AOWNER | XFS_ILOG_DOWNER);
2317 spin_unlock(&iip->ili_lock);
2318
2319 /*
2320 * Bump the generation count so no one will be confused
2321 * by reincarnations of this inode.
2322 */
2323 VFS_I(ip)->i_generation++;
2324 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
2325
2326 if (xic.deleted)
2327 error = xfs_ifree_cluster(tp, pag, ip, &xic);
2328 out:
2329 xfs_perag_put(pag);
2330 return error;
2331 }
2332
2333 /*
2334 * This is called to unpin an inode. The caller must have the inode locked
2335 * in at least shared mode so that the buffer cannot be subsequently pinned
2336 * once someone is waiting for it to be unpinned.
2337 */
2338 static void
xfs_iunpin(struct xfs_inode * ip)2339 xfs_iunpin(
2340 struct xfs_inode *ip)
2341 {
2342 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
2343
2344 trace_xfs_inode_unpin_nowait(ip, _RET_IP_);
2345
2346 /* Give the log a push to start the unpinning I/O */
2347 xfs_log_force_seq(ip->i_mount, ip->i_itemp->ili_commit_seq, 0, NULL);
2348
2349 }
2350
2351 static void
__xfs_iunpin_wait(struct xfs_inode * ip)2352 __xfs_iunpin_wait(
2353 struct xfs_inode *ip)
2354 {
2355 wait_queue_head_t *wq = bit_waitqueue(&ip->i_flags, __XFS_IPINNED_BIT);
2356 DEFINE_WAIT_BIT(wait, &ip->i_flags, __XFS_IPINNED_BIT);
2357
2358 xfs_iunpin(ip);
2359
2360 do {
2361 prepare_to_wait(wq, &wait.wq_entry, TASK_UNINTERRUPTIBLE);
2362 if (xfs_ipincount(ip))
2363 io_schedule();
2364 } while (xfs_ipincount(ip));
2365 finish_wait(wq, &wait.wq_entry);
2366 }
2367
2368 void
xfs_iunpin_wait(struct xfs_inode * ip)2369 xfs_iunpin_wait(
2370 struct xfs_inode *ip)
2371 {
2372 if (xfs_ipincount(ip))
2373 __xfs_iunpin_wait(ip);
2374 }
2375
2376 /*
2377 * Removing an inode from the namespace involves removing the directory entry
2378 * and dropping the link count on the inode. Removing the directory entry can
2379 * result in locking an AGF (directory blocks were freed) and removing a link
2380 * count can result in placing the inode on an unlinked list which results in
2381 * locking an AGI.
2382 *
2383 * The big problem here is that we have an ordering constraint on AGF and AGI
2384 * locking - inode allocation locks the AGI, then can allocate a new extent for
2385 * new inodes, locking the AGF after the AGI. Similarly, freeing the inode
2386 * removes the inode from the unlinked list, requiring that we lock the AGI
2387 * first, and then freeing the inode can result in an inode chunk being freed
2388 * and hence freeing disk space requiring that we lock an AGF.
2389 *
2390 * Hence the ordering that is imposed by other parts of the code is AGI before
2391 * AGF. This means we cannot remove the directory entry before we drop the inode
2392 * reference count and put it on the unlinked list as this results in a lock
2393 * order of AGF then AGI, and this can deadlock against inode allocation and
2394 * freeing. Therefore we must drop the link counts before we remove the
2395 * directory entry.
2396 *
2397 * This is still safe from a transactional point of view - it is not until we
2398 * get to xfs_defer_finish() that we have the possibility of multiple
2399 * transactions in this operation. Hence as long as we remove the directory
2400 * entry and drop the link count in the first transaction of the remove
2401 * operation, there are no transactional constraints on the ordering here.
2402 */
2403 int
xfs_remove(xfs_inode_t * dp,struct xfs_name * name,xfs_inode_t * ip)2404 xfs_remove(
2405 xfs_inode_t *dp,
2406 struct xfs_name *name,
2407 xfs_inode_t *ip)
2408 {
2409 xfs_mount_t *mp = dp->i_mount;
2410 xfs_trans_t *tp = NULL;
2411 int is_dir = S_ISDIR(VFS_I(ip)->i_mode);
2412 int dontcare;
2413 int error = 0;
2414 uint resblks;
2415
2416 trace_xfs_remove(dp, name);
2417
2418 if (xfs_is_shutdown(mp))
2419 return -EIO;
2420
2421 error = xfs_qm_dqattach(dp);
2422 if (error)
2423 goto std_return;
2424
2425 error = xfs_qm_dqattach(ip);
2426 if (error)
2427 goto std_return;
2428
2429 /*
2430 * We try to get the real space reservation first, allowing for
2431 * directory btree deletion(s) implying possible bmap insert(s). If we
2432 * can't get the space reservation then we use 0 instead, and avoid the
2433 * bmap btree insert(s) in the directory code by, if the bmap insert
2434 * tries to happen, instead trimming the LAST block from the directory.
2435 *
2436 * Ignore EDQUOT and ENOSPC being returned via nospace_error because
2437 * the directory code can handle a reservationless update and we don't
2438 * want to prevent a user from trying to free space by deleting things.
2439 */
2440 resblks = XFS_REMOVE_SPACE_RES(mp);
2441 error = xfs_trans_alloc_dir(dp, &M_RES(mp)->tr_remove, ip, &resblks,
2442 &tp, &dontcare);
2443 if (error) {
2444 ASSERT(error != -ENOSPC);
2445 goto std_return;
2446 }
2447
2448 /*
2449 * If we're removing a directory perform some additional validation.
2450 */
2451 if (is_dir) {
2452 ASSERT(VFS_I(ip)->i_nlink >= 2);
2453 if (VFS_I(ip)->i_nlink != 2) {
2454 error = -ENOTEMPTY;
2455 goto out_trans_cancel;
2456 }
2457 if (!xfs_dir_isempty(ip)) {
2458 error = -ENOTEMPTY;
2459 goto out_trans_cancel;
2460 }
2461
2462 /* Drop the link from ip's "..". */
2463 error = xfs_droplink(tp, dp);
2464 if (error)
2465 goto out_trans_cancel;
2466
2467 /* Drop the "." link from ip to self. */
2468 error = xfs_droplink(tp, ip);
2469 if (error)
2470 goto out_trans_cancel;
2471
2472 /*
2473 * Point the unlinked child directory's ".." entry to the root
2474 * directory to eliminate back-references to inodes that may
2475 * get freed before the child directory is closed. If the fs
2476 * gets shrunk, this can lead to dirent inode validation errors.
2477 */
2478 if (dp->i_ino != tp->t_mountp->m_sb.sb_rootino) {
2479 error = xfs_dir_replace(tp, ip, &xfs_name_dotdot,
2480 tp->t_mountp->m_sb.sb_rootino, 0);
2481 if (error)
2482 return error;
2483 }
2484 } else {
2485 /*
2486 * When removing a non-directory we need to log the parent
2487 * inode here. For a directory this is done implicitly
2488 * by the xfs_droplink call for the ".." entry.
2489 */
2490 xfs_trans_log_inode(tp, dp, XFS_ILOG_CORE);
2491 }
2492 xfs_trans_ichgtime(tp, dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2493
2494 /* Drop the link from dp to ip. */
2495 error = xfs_droplink(tp, ip);
2496 if (error)
2497 goto out_trans_cancel;
2498
2499 error = xfs_dir_removename(tp, dp, name, ip->i_ino, resblks);
2500 if (error) {
2501 ASSERT(error != -ENOENT);
2502 goto out_trans_cancel;
2503 }
2504
2505 /*
2506 * If this is a synchronous mount, make sure that the
2507 * remove transaction goes to disk before returning to
2508 * the user.
2509 */
2510 if (xfs_has_wsync(mp) || xfs_has_dirsync(mp))
2511 xfs_trans_set_sync(tp);
2512
2513 error = xfs_trans_commit(tp);
2514 if (error)
2515 goto std_return;
2516
2517 if (is_dir && xfs_inode_is_filestream(ip))
2518 xfs_filestream_deassociate(ip);
2519
2520 return 0;
2521
2522 out_trans_cancel:
2523 xfs_trans_cancel(tp);
2524 std_return:
2525 return error;
2526 }
2527
2528 /*
2529 * Enter all inodes for a rename transaction into a sorted array.
2530 */
2531 #define __XFS_SORT_INODES 5
2532 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)2533 xfs_sort_for_rename(
2534 struct xfs_inode *dp1, /* in: old (source) directory inode */
2535 struct xfs_inode *dp2, /* in: new (target) directory inode */
2536 struct xfs_inode *ip1, /* in: inode of old entry */
2537 struct xfs_inode *ip2, /* in: inode of new entry */
2538 struct xfs_inode *wip, /* in: whiteout inode */
2539 struct xfs_inode **i_tab,/* out: sorted array of inodes */
2540 int *num_inodes) /* in/out: inodes in array */
2541 {
2542 int i, j;
2543
2544 ASSERT(*num_inodes == __XFS_SORT_INODES);
2545 memset(i_tab, 0, *num_inodes * sizeof(struct xfs_inode *));
2546
2547 /*
2548 * i_tab contains a list of pointers to inodes. We initialize
2549 * the table here & we'll sort it. We will then use it to
2550 * order the acquisition of the inode locks.
2551 *
2552 * Note that the table may contain duplicates. e.g., dp1 == dp2.
2553 */
2554 i = 0;
2555 i_tab[i++] = dp1;
2556 i_tab[i++] = dp2;
2557 i_tab[i++] = ip1;
2558 if (ip2)
2559 i_tab[i++] = ip2;
2560 if (wip)
2561 i_tab[i++] = wip;
2562 *num_inodes = i;
2563
2564 /*
2565 * Sort the elements via bubble sort. (Remember, there are at
2566 * most 5 elements to sort, so this is adequate.)
2567 */
2568 for (i = 0; i < *num_inodes; i++) {
2569 for (j = 1; j < *num_inodes; j++) {
2570 if (i_tab[j]->i_ino < i_tab[j-1]->i_ino) {
2571 struct xfs_inode *temp = i_tab[j];
2572 i_tab[j] = i_tab[j-1];
2573 i_tab[j-1] = temp;
2574 }
2575 }
2576 }
2577 }
2578
2579 static int
xfs_finish_rename(struct xfs_trans * tp)2580 xfs_finish_rename(
2581 struct xfs_trans *tp)
2582 {
2583 /*
2584 * If this is a synchronous mount, make sure that the rename transaction
2585 * goes to disk before returning to the user.
2586 */
2587 if (xfs_has_wsync(tp->t_mountp) || xfs_has_dirsync(tp->t_mountp))
2588 xfs_trans_set_sync(tp);
2589
2590 return xfs_trans_commit(tp);
2591 }
2592
2593 /*
2594 * xfs_cross_rename()
2595 *
2596 * responsible for handling RENAME_EXCHANGE flag in renameat2() syscall
2597 */
2598 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)2599 xfs_cross_rename(
2600 struct xfs_trans *tp,
2601 struct xfs_inode *dp1,
2602 struct xfs_name *name1,
2603 struct xfs_inode *ip1,
2604 struct xfs_inode *dp2,
2605 struct xfs_name *name2,
2606 struct xfs_inode *ip2,
2607 int spaceres)
2608 {
2609 int error = 0;
2610 int ip1_flags = 0;
2611 int ip2_flags = 0;
2612 int dp2_flags = 0;
2613
2614 /* Swap inode number for dirent in first parent */
2615 error = xfs_dir_replace(tp, dp1, name1, ip2->i_ino, spaceres);
2616 if (error)
2617 goto out_trans_abort;
2618
2619 /* Swap inode number for dirent in second parent */
2620 error = xfs_dir_replace(tp, dp2, name2, ip1->i_ino, spaceres);
2621 if (error)
2622 goto out_trans_abort;
2623
2624 /*
2625 * If we're renaming one or more directories across different parents,
2626 * update the respective ".." entries (and link counts) to match the new
2627 * parents.
2628 */
2629 if (dp1 != dp2) {
2630 dp2_flags = XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2631
2632 if (S_ISDIR(VFS_I(ip2)->i_mode)) {
2633 error = xfs_dir_replace(tp, ip2, &xfs_name_dotdot,
2634 dp1->i_ino, spaceres);
2635 if (error)
2636 goto out_trans_abort;
2637
2638 /* transfer ip2 ".." reference to dp1 */
2639 if (!S_ISDIR(VFS_I(ip1)->i_mode)) {
2640 error = xfs_droplink(tp, dp2);
2641 if (error)
2642 goto out_trans_abort;
2643 xfs_bumplink(tp, dp1);
2644 }
2645
2646 /*
2647 * Although ip1 isn't changed here, userspace needs
2648 * to be warned about the change, so that applications
2649 * relying on it (like backup ones), will properly
2650 * notify the change
2651 */
2652 ip1_flags |= XFS_ICHGTIME_CHG;
2653 ip2_flags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2654 }
2655
2656 if (S_ISDIR(VFS_I(ip1)->i_mode)) {
2657 error = xfs_dir_replace(tp, ip1, &xfs_name_dotdot,
2658 dp2->i_ino, spaceres);
2659 if (error)
2660 goto out_trans_abort;
2661
2662 /* transfer ip1 ".." reference to dp2 */
2663 if (!S_ISDIR(VFS_I(ip2)->i_mode)) {
2664 error = xfs_droplink(tp, dp1);
2665 if (error)
2666 goto out_trans_abort;
2667 xfs_bumplink(tp, dp2);
2668 }
2669
2670 /*
2671 * Although ip2 isn't changed here, userspace needs
2672 * to be warned about the change, so that applications
2673 * relying on it (like backup ones), will properly
2674 * notify the change
2675 */
2676 ip1_flags |= XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG;
2677 ip2_flags |= XFS_ICHGTIME_CHG;
2678 }
2679 }
2680
2681 if (ip1_flags) {
2682 xfs_trans_ichgtime(tp, ip1, ip1_flags);
2683 xfs_trans_log_inode(tp, ip1, XFS_ILOG_CORE);
2684 }
2685 if (ip2_flags) {
2686 xfs_trans_ichgtime(tp, ip2, ip2_flags);
2687 xfs_trans_log_inode(tp, ip2, XFS_ILOG_CORE);
2688 }
2689 if (dp2_flags) {
2690 xfs_trans_ichgtime(tp, dp2, dp2_flags);
2691 xfs_trans_log_inode(tp, dp2, XFS_ILOG_CORE);
2692 }
2693 xfs_trans_ichgtime(tp, dp1, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2694 xfs_trans_log_inode(tp, dp1, XFS_ILOG_CORE);
2695 return xfs_finish_rename(tp);
2696
2697 out_trans_abort:
2698 xfs_trans_cancel(tp);
2699 return error;
2700 }
2701
2702 /*
2703 * xfs_rename_alloc_whiteout()
2704 *
2705 * Return a referenced, unlinked, unlocked inode that can be used as a
2706 * whiteout in a rename transaction. We use a tmpfile inode here so that if we
2707 * crash between allocating the inode and linking it into the rename transaction
2708 * recovery will free the inode and we won't leak it.
2709 */
2710 static int
xfs_rename_alloc_whiteout(struct user_namespace * mnt_userns,struct xfs_name * src_name,struct xfs_inode * dp,struct xfs_inode ** wip)2711 xfs_rename_alloc_whiteout(
2712 struct user_namespace *mnt_userns,
2713 struct xfs_name *src_name,
2714 struct xfs_inode *dp,
2715 struct xfs_inode **wip)
2716 {
2717 struct xfs_inode *tmpfile;
2718 struct qstr name;
2719 int error;
2720
2721 error = xfs_create_tmpfile(mnt_userns, dp, S_IFCHR | WHITEOUT_MODE,
2722 &tmpfile);
2723 if (error)
2724 return error;
2725
2726 name.name = src_name->name;
2727 name.len = src_name->len;
2728 error = xfs_inode_init_security(VFS_I(tmpfile), VFS_I(dp), &name);
2729 if (error) {
2730 xfs_finish_inode_setup(tmpfile);
2731 xfs_irele(tmpfile);
2732 return error;
2733 }
2734
2735 /*
2736 * Prepare the tmpfile inode as if it were created through the VFS.
2737 * Complete the inode setup and flag it as linkable. nlink is already
2738 * zero, so we can skip the drop_nlink.
2739 */
2740 xfs_setup_iops(tmpfile);
2741 xfs_finish_inode_setup(tmpfile);
2742 VFS_I(tmpfile)->i_state |= I_LINKABLE;
2743
2744 *wip = tmpfile;
2745 return 0;
2746 }
2747
2748 /*
2749 * xfs_rename
2750 */
2751 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)2752 xfs_rename(
2753 struct user_namespace *mnt_userns,
2754 struct xfs_inode *src_dp,
2755 struct xfs_name *src_name,
2756 struct xfs_inode *src_ip,
2757 struct xfs_inode *target_dp,
2758 struct xfs_name *target_name,
2759 struct xfs_inode *target_ip,
2760 unsigned int flags)
2761 {
2762 struct xfs_mount *mp = src_dp->i_mount;
2763 struct xfs_trans *tp;
2764 struct xfs_inode *wip = NULL; /* whiteout inode */
2765 struct xfs_inode *inodes[__XFS_SORT_INODES];
2766 int i;
2767 int num_inodes = __XFS_SORT_INODES;
2768 bool new_parent = (src_dp != target_dp);
2769 bool src_is_directory = S_ISDIR(VFS_I(src_ip)->i_mode);
2770 int spaceres;
2771 bool retried = false;
2772 int error, nospace_error = 0;
2773
2774 trace_xfs_rename(src_dp, target_dp, src_name, target_name);
2775
2776 if ((flags & RENAME_EXCHANGE) && !target_ip)
2777 return -EINVAL;
2778
2779 /*
2780 * If we are doing a whiteout operation, allocate the whiteout inode
2781 * we will be placing at the target and ensure the type is set
2782 * appropriately.
2783 */
2784 if (flags & RENAME_WHITEOUT) {
2785 error = xfs_rename_alloc_whiteout(mnt_userns, src_name,
2786 target_dp, &wip);
2787 if (error)
2788 return error;
2789
2790 /* setup target dirent info as whiteout */
2791 src_name->type = XFS_DIR3_FT_CHRDEV;
2792 }
2793
2794 xfs_sort_for_rename(src_dp, target_dp, src_ip, target_ip, wip,
2795 inodes, &num_inodes);
2796
2797 retry:
2798 nospace_error = 0;
2799 spaceres = XFS_RENAME_SPACE_RES(mp, target_name->len);
2800 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, spaceres, 0, 0, &tp);
2801 if (error == -ENOSPC) {
2802 nospace_error = error;
2803 spaceres = 0;
2804 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_rename, 0, 0, 0,
2805 &tp);
2806 }
2807 if (error)
2808 goto out_release_wip;
2809
2810 /*
2811 * Attach the dquots to the inodes
2812 */
2813 error = xfs_qm_vop_rename_dqattach(inodes);
2814 if (error)
2815 goto out_trans_cancel;
2816
2817 /*
2818 * Lock all the participating inodes. Depending upon whether
2819 * the target_name exists in the target directory, and
2820 * whether the target directory is the same as the source
2821 * directory, we can lock from 2 to 5 inodes.
2822 */
2823 xfs_lock_inodes(inodes, num_inodes, XFS_ILOCK_EXCL);
2824
2825 /*
2826 * Join all the inodes to the transaction. From this point on,
2827 * we can rely on either trans_commit or trans_cancel to unlock
2828 * them.
2829 */
2830 xfs_trans_ijoin(tp, src_dp, XFS_ILOCK_EXCL);
2831 if (new_parent)
2832 xfs_trans_ijoin(tp, target_dp, XFS_ILOCK_EXCL);
2833 xfs_trans_ijoin(tp, src_ip, XFS_ILOCK_EXCL);
2834 if (target_ip)
2835 xfs_trans_ijoin(tp, target_ip, XFS_ILOCK_EXCL);
2836 if (wip)
2837 xfs_trans_ijoin(tp, wip, XFS_ILOCK_EXCL);
2838
2839 /*
2840 * If we are using project inheritance, we only allow renames
2841 * into our tree when the project IDs are the same; else the
2842 * tree quota mechanism would be circumvented.
2843 */
2844 if (unlikely((target_dp->i_diflags & XFS_DIFLAG_PROJINHERIT) &&
2845 target_dp->i_projid != src_ip->i_projid)) {
2846 error = -EXDEV;
2847 goto out_trans_cancel;
2848 }
2849
2850 /* RENAME_EXCHANGE is unique from here on. */
2851 if (flags & RENAME_EXCHANGE)
2852 return xfs_cross_rename(tp, src_dp, src_name, src_ip,
2853 target_dp, target_name, target_ip,
2854 spaceres);
2855
2856 /*
2857 * Try to reserve quota to handle an expansion of the target directory.
2858 * We'll allow the rename to continue in reservationless mode if we hit
2859 * a space usage constraint. If we trigger reservationless mode, save
2860 * the errno if there isn't any free space in the target directory.
2861 */
2862 if (spaceres != 0) {
2863 error = xfs_trans_reserve_quota_nblks(tp, target_dp, spaceres,
2864 0, false);
2865 if (error == -EDQUOT || error == -ENOSPC) {
2866 if (!retried) {
2867 xfs_trans_cancel(tp);
2868 xfs_blockgc_free_quota(target_dp, 0);
2869 retried = true;
2870 goto retry;
2871 }
2872
2873 nospace_error = error;
2874 spaceres = 0;
2875 error = 0;
2876 }
2877 if (error)
2878 goto out_trans_cancel;
2879 }
2880
2881 /*
2882 * Check for expected errors before we dirty the transaction
2883 * so we can return an error without a transaction abort.
2884 */
2885 if (target_ip == NULL) {
2886 /*
2887 * If there's no space reservation, check the entry will
2888 * fit before actually inserting it.
2889 */
2890 if (!spaceres) {
2891 error = xfs_dir_canenter(tp, target_dp, target_name);
2892 if (error)
2893 goto out_trans_cancel;
2894 }
2895 } else {
2896 /*
2897 * If target exists and it's a directory, check that whether
2898 * it can be destroyed.
2899 */
2900 if (S_ISDIR(VFS_I(target_ip)->i_mode) &&
2901 (!xfs_dir_isempty(target_ip) ||
2902 (VFS_I(target_ip)->i_nlink > 2))) {
2903 error = -EEXIST;
2904 goto out_trans_cancel;
2905 }
2906 }
2907
2908 /*
2909 * Lock the AGI buffers we need to handle bumping the nlink of the
2910 * whiteout inode off the unlinked list and to handle dropping the
2911 * nlink of the target inode. Per locking order rules, do this in
2912 * increasing AG order and before directory block allocation tries to
2913 * grab AGFs because we grab AGIs before AGFs.
2914 *
2915 * The (vfs) caller must ensure that if src is a directory then
2916 * target_ip is either null or an empty directory.
2917 */
2918 for (i = 0; i < num_inodes && inodes[i] != NULL; i++) {
2919 if (inodes[i] == wip ||
2920 (inodes[i] == target_ip &&
2921 (VFS_I(target_ip)->i_nlink == 1 || src_is_directory))) {
2922 struct xfs_perag *pag;
2923 struct xfs_buf *bp;
2924
2925 pag = xfs_perag_get(mp,
2926 XFS_INO_TO_AGNO(mp, inodes[i]->i_ino));
2927 error = xfs_read_agi(pag, tp, &bp);
2928 xfs_perag_put(pag);
2929 if (error)
2930 goto out_trans_cancel;
2931 }
2932 }
2933
2934 /*
2935 * Directory entry creation below may acquire the AGF. Remove
2936 * the whiteout from the unlinked list first to preserve correct
2937 * AGI/AGF locking order. This dirties the transaction so failures
2938 * after this point will abort and log recovery will clean up the
2939 * mess.
2940 *
2941 * For whiteouts, we need to bump the link count on the whiteout
2942 * inode. After this point, we have a real link, clear the tmpfile
2943 * state flag from the inode so it doesn't accidentally get misused
2944 * in future.
2945 */
2946 if (wip) {
2947 struct xfs_perag *pag;
2948
2949 ASSERT(VFS_I(wip)->i_nlink == 0);
2950
2951 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, wip->i_ino));
2952 error = xfs_iunlink_remove(tp, pag, wip);
2953 xfs_perag_put(pag);
2954 if (error)
2955 goto out_trans_cancel;
2956
2957 xfs_bumplink(tp, wip);
2958 VFS_I(wip)->i_state &= ~I_LINKABLE;
2959 }
2960
2961 /*
2962 * Set up the target.
2963 */
2964 if (target_ip == NULL) {
2965 /*
2966 * If target does not exist and the rename crosses
2967 * directories, adjust the target directory link count
2968 * to account for the ".." reference from the new entry.
2969 */
2970 error = xfs_dir_createname(tp, target_dp, target_name,
2971 src_ip->i_ino, spaceres);
2972 if (error)
2973 goto out_trans_cancel;
2974
2975 xfs_trans_ichgtime(tp, target_dp,
2976 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2977
2978 if (new_parent && src_is_directory) {
2979 xfs_bumplink(tp, target_dp);
2980 }
2981 } else { /* target_ip != NULL */
2982 /*
2983 * Link the source inode under the target name.
2984 * If the source inode is a directory and we are moving
2985 * it across directories, its ".." entry will be
2986 * inconsistent until we replace that down below.
2987 *
2988 * In case there is already an entry with the same
2989 * name at the destination directory, remove it first.
2990 */
2991 error = xfs_dir_replace(tp, target_dp, target_name,
2992 src_ip->i_ino, spaceres);
2993 if (error)
2994 goto out_trans_cancel;
2995
2996 xfs_trans_ichgtime(tp, target_dp,
2997 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
2998
2999 /*
3000 * Decrement the link count on the target since the target
3001 * dir no longer points to it.
3002 */
3003 error = xfs_droplink(tp, target_ip);
3004 if (error)
3005 goto out_trans_cancel;
3006
3007 if (src_is_directory) {
3008 /*
3009 * Drop the link from the old "." entry.
3010 */
3011 error = xfs_droplink(tp, target_ip);
3012 if (error)
3013 goto out_trans_cancel;
3014 }
3015 } /* target_ip != NULL */
3016
3017 /*
3018 * Remove the source.
3019 */
3020 if (new_parent && src_is_directory) {
3021 /*
3022 * Rewrite the ".." entry to point to the new
3023 * directory.
3024 */
3025 error = xfs_dir_replace(tp, src_ip, &xfs_name_dotdot,
3026 target_dp->i_ino, spaceres);
3027 ASSERT(error != -EEXIST);
3028 if (error)
3029 goto out_trans_cancel;
3030 }
3031
3032 /*
3033 * We always want to hit the ctime on the source inode.
3034 *
3035 * This isn't strictly required by the standards since the source
3036 * inode isn't really being changed, but old unix file systems did
3037 * it and some incremental backup programs won't work without it.
3038 */
3039 xfs_trans_ichgtime(tp, src_ip, XFS_ICHGTIME_CHG);
3040 xfs_trans_log_inode(tp, src_ip, XFS_ILOG_CORE);
3041
3042 /*
3043 * Adjust the link count on src_dp. This is necessary when
3044 * renaming a directory, either within one parent when
3045 * the target existed, or across two parent directories.
3046 */
3047 if (src_is_directory && (new_parent || target_ip != NULL)) {
3048
3049 /*
3050 * Decrement link count on src_directory since the
3051 * entry that's moved no longer points to it.
3052 */
3053 error = xfs_droplink(tp, src_dp);
3054 if (error)
3055 goto out_trans_cancel;
3056 }
3057
3058 /*
3059 * For whiteouts, we only need to update the source dirent with the
3060 * inode number of the whiteout inode rather than removing it
3061 * altogether.
3062 */
3063 if (wip)
3064 error = xfs_dir_replace(tp, src_dp, src_name, wip->i_ino,
3065 spaceres);
3066 else
3067 error = xfs_dir_removename(tp, src_dp, src_name, src_ip->i_ino,
3068 spaceres);
3069
3070 if (error)
3071 goto out_trans_cancel;
3072
3073 xfs_trans_ichgtime(tp, src_dp, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
3074 xfs_trans_log_inode(tp, src_dp, XFS_ILOG_CORE);
3075 if (new_parent)
3076 xfs_trans_log_inode(tp, target_dp, XFS_ILOG_CORE);
3077
3078 error = xfs_finish_rename(tp);
3079 if (wip)
3080 xfs_irele(wip);
3081 return error;
3082
3083 out_trans_cancel:
3084 xfs_trans_cancel(tp);
3085 out_release_wip:
3086 if (wip)
3087 xfs_irele(wip);
3088 if (error == -ENOSPC && nospace_error)
3089 error = nospace_error;
3090 return error;
3091 }
3092
3093 static int
xfs_iflush(struct xfs_inode * ip,struct xfs_buf * bp)3094 xfs_iflush(
3095 struct xfs_inode *ip,
3096 struct xfs_buf *bp)
3097 {
3098 struct xfs_inode_log_item *iip = ip->i_itemp;
3099 struct xfs_dinode *dip;
3100 struct xfs_mount *mp = ip->i_mount;
3101 int error;
3102
3103 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL|XFS_ILOCK_SHARED));
3104 ASSERT(xfs_iflags_test(ip, XFS_IFLUSHING));
3105 ASSERT(ip->i_df.if_format != XFS_DINODE_FMT_BTREE ||
3106 ip->i_df.if_nextents > XFS_IFORK_MAXEXT(ip, XFS_DATA_FORK));
3107 ASSERT(iip->ili_item.li_buf == bp);
3108
3109 dip = xfs_buf_offset(bp, ip->i_imap.im_boffset);
3110
3111 /*
3112 * We don't flush the inode if any of the following checks fail, but we
3113 * do still update the log item and attach to the backing buffer as if
3114 * the flush happened. This is a formality to facilitate predictable
3115 * error handling as the caller will shutdown and fail the buffer.
3116 */
3117 error = -EFSCORRUPTED;
3118 if (XFS_TEST_ERROR(dip->di_magic != cpu_to_be16(XFS_DINODE_MAGIC),
3119 mp, XFS_ERRTAG_IFLUSH_1)) {
3120 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3121 "%s: Bad inode %llu magic number 0x%x, ptr "PTR_FMT,
3122 __func__, ip->i_ino, be16_to_cpu(dip->di_magic), dip);
3123 goto flush_out;
3124 }
3125 if (S_ISREG(VFS_I(ip)->i_mode)) {
3126 if (XFS_TEST_ERROR(
3127 ip->i_df.if_format != XFS_DINODE_FMT_EXTENTS &&
3128 ip->i_df.if_format != XFS_DINODE_FMT_BTREE,
3129 mp, XFS_ERRTAG_IFLUSH_3)) {
3130 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3131 "%s: Bad regular inode %llu, ptr "PTR_FMT,
3132 __func__, ip->i_ino, ip);
3133 goto flush_out;
3134 }
3135 } else if (S_ISDIR(VFS_I(ip)->i_mode)) {
3136 if (XFS_TEST_ERROR(
3137 ip->i_df.if_format != XFS_DINODE_FMT_EXTENTS &&
3138 ip->i_df.if_format != XFS_DINODE_FMT_BTREE &&
3139 ip->i_df.if_format != XFS_DINODE_FMT_LOCAL,
3140 mp, XFS_ERRTAG_IFLUSH_4)) {
3141 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3142 "%s: Bad directory inode %llu, ptr "PTR_FMT,
3143 __func__, ip->i_ino, ip);
3144 goto flush_out;
3145 }
3146 }
3147 if (XFS_TEST_ERROR(ip->i_df.if_nextents + xfs_ifork_nextents(&ip->i_af) >
3148 ip->i_nblocks, mp, XFS_ERRTAG_IFLUSH_5)) {
3149 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3150 "%s: detected corrupt incore inode %llu, "
3151 "total extents = %llu nblocks = %lld, ptr "PTR_FMT,
3152 __func__, ip->i_ino,
3153 ip->i_df.if_nextents + xfs_ifork_nextents(&ip->i_af),
3154 ip->i_nblocks, ip);
3155 goto flush_out;
3156 }
3157 if (XFS_TEST_ERROR(ip->i_forkoff > mp->m_sb.sb_inodesize,
3158 mp, XFS_ERRTAG_IFLUSH_6)) {
3159 xfs_alert_tag(mp, XFS_PTAG_IFLUSH,
3160 "%s: bad inode %llu, forkoff 0x%x, ptr "PTR_FMT,
3161 __func__, ip->i_ino, ip->i_forkoff, ip);
3162 goto flush_out;
3163 }
3164
3165 /*
3166 * Inode item log recovery for v2 inodes are dependent on the flushiter
3167 * count for correct sequencing. We bump the flush iteration count so
3168 * we can detect flushes which postdate a log record during recovery.
3169 * This is redundant as we now log every change and hence this can't
3170 * happen but we need to still do it to ensure backwards compatibility
3171 * with old kernels that predate logging all inode changes.
3172 */
3173 if (!xfs_has_v3inodes(mp))
3174 ip->i_flushiter++;
3175
3176 /*
3177 * If there are inline format data / attr forks attached to this inode,
3178 * make sure they are not corrupt.
3179 */
3180 if (ip->i_df.if_format == XFS_DINODE_FMT_LOCAL &&
3181 xfs_ifork_verify_local_data(ip))
3182 goto flush_out;
3183 if (xfs_inode_has_attr_fork(ip) &&
3184 ip->i_af.if_format == XFS_DINODE_FMT_LOCAL &&
3185 xfs_ifork_verify_local_attr(ip))
3186 goto flush_out;
3187
3188 /*
3189 * Copy the dirty parts of the inode into the on-disk inode. We always
3190 * copy out the core of the inode, because if the inode is dirty at all
3191 * the core must be.
3192 */
3193 xfs_inode_to_disk(ip, dip, iip->ili_item.li_lsn);
3194
3195 /* Wrap, we never let the log put out DI_MAX_FLUSH */
3196 if (!xfs_has_v3inodes(mp)) {
3197 if (ip->i_flushiter == DI_MAX_FLUSH)
3198 ip->i_flushiter = 0;
3199 }
3200
3201 xfs_iflush_fork(ip, dip, iip, XFS_DATA_FORK);
3202 if (xfs_inode_has_attr_fork(ip))
3203 xfs_iflush_fork(ip, dip, iip, XFS_ATTR_FORK);
3204
3205 /*
3206 * We've recorded everything logged in the inode, so we'd like to clear
3207 * the ili_fields bits so we don't log and flush things unnecessarily.
3208 * However, we can't stop logging all this information until the data
3209 * we've copied into the disk buffer is written to disk. If we did we
3210 * might overwrite the copy of the inode in the log with all the data
3211 * after re-logging only part of it, and in the face of a crash we
3212 * wouldn't have all the data we need to recover.
3213 *
3214 * What we do is move the bits to the ili_last_fields field. When
3215 * logging the inode, these bits are moved back to the ili_fields field.
3216 * In the xfs_buf_inode_iodone() routine we clear ili_last_fields, since
3217 * we know that the information those bits represent is permanently on
3218 * disk. As long as the flush completes before the inode is logged
3219 * again, then both ili_fields and ili_last_fields will be cleared.
3220 */
3221 error = 0;
3222 flush_out:
3223 spin_lock(&iip->ili_lock);
3224 iip->ili_last_fields = iip->ili_fields;
3225 iip->ili_fields = 0;
3226 iip->ili_fsync_fields = 0;
3227 spin_unlock(&iip->ili_lock);
3228
3229 /*
3230 * Store the current LSN of the inode so that we can tell whether the
3231 * item has moved in the AIL from xfs_buf_inode_iodone().
3232 */
3233 xfs_trans_ail_copy_lsn(mp->m_ail, &iip->ili_flush_lsn,
3234 &iip->ili_item.li_lsn);
3235
3236 /* generate the checksum. */
3237 xfs_dinode_calc_crc(mp, dip);
3238 return error;
3239 }
3240
3241 /*
3242 * Non-blocking flush of dirty inode metadata into the backing buffer.
3243 *
3244 * The caller must have a reference to the inode and hold the cluster buffer
3245 * locked. The function will walk across all the inodes on the cluster buffer it
3246 * can find and lock without blocking, and flush them to the cluster buffer.
3247 *
3248 * On successful flushing of at least one inode, the caller must write out the
3249 * buffer and release it. If no inodes are flushed, -EAGAIN will be returned and
3250 * the caller needs to release the buffer. On failure, the filesystem will be
3251 * shut down, the buffer will have been unlocked and released, and EFSCORRUPTED
3252 * will be returned.
3253 */
3254 int
xfs_iflush_cluster(struct xfs_buf * bp)3255 xfs_iflush_cluster(
3256 struct xfs_buf *bp)
3257 {
3258 struct xfs_mount *mp = bp->b_mount;
3259 struct xfs_log_item *lip, *n;
3260 struct xfs_inode *ip;
3261 struct xfs_inode_log_item *iip;
3262 int clcount = 0;
3263 int error = 0;
3264
3265 /*
3266 * We must use the safe variant here as on shutdown xfs_iflush_abort()
3267 * will remove itself from the list.
3268 */
3269 list_for_each_entry_safe(lip, n, &bp->b_li_list, li_bio_list) {
3270 iip = (struct xfs_inode_log_item *)lip;
3271 ip = iip->ili_inode;
3272
3273 /*
3274 * Quick and dirty check to avoid locks if possible.
3275 */
3276 if (__xfs_iflags_test(ip, XFS_IRECLAIM | XFS_IFLUSHING))
3277 continue;
3278 if (xfs_ipincount(ip))
3279 continue;
3280
3281 /*
3282 * The inode is still attached to the buffer, which means it is
3283 * dirty but reclaim might try to grab it. Check carefully for
3284 * that, and grab the ilock while still holding the i_flags_lock
3285 * to guarantee reclaim will not be able to reclaim this inode
3286 * once we drop the i_flags_lock.
3287 */
3288 spin_lock(&ip->i_flags_lock);
3289 ASSERT(!__xfs_iflags_test(ip, XFS_ISTALE));
3290 if (__xfs_iflags_test(ip, XFS_IRECLAIM | XFS_IFLUSHING)) {
3291 spin_unlock(&ip->i_flags_lock);
3292 continue;
3293 }
3294
3295 /*
3296 * ILOCK will pin the inode against reclaim and prevent
3297 * concurrent transactions modifying the inode while we are
3298 * flushing the inode. If we get the lock, set the flushing
3299 * state before we drop the i_flags_lock.
3300 */
3301 if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
3302 spin_unlock(&ip->i_flags_lock);
3303 continue;
3304 }
3305 __xfs_iflags_set(ip, XFS_IFLUSHING);
3306 spin_unlock(&ip->i_flags_lock);
3307
3308 /*
3309 * Abort flushing this inode if we are shut down because the
3310 * inode may not currently be in the AIL. This can occur when
3311 * log I/O failure unpins the inode without inserting into the
3312 * AIL, leaving a dirty/unpinned inode attached to the buffer
3313 * that otherwise looks like it should be flushed.
3314 */
3315 if (xlog_is_shutdown(mp->m_log)) {
3316 xfs_iunpin_wait(ip);
3317 xfs_iflush_abort(ip);
3318 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3319 error = -EIO;
3320 continue;
3321 }
3322
3323 /* don't block waiting on a log force to unpin dirty inodes */
3324 if (xfs_ipincount(ip)) {
3325 xfs_iflags_clear(ip, XFS_IFLUSHING);
3326 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3327 continue;
3328 }
3329
3330 if (!xfs_inode_clean(ip))
3331 error = xfs_iflush(ip, bp);
3332 else
3333 xfs_iflags_clear(ip, XFS_IFLUSHING);
3334 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3335 if (error)
3336 break;
3337 clcount++;
3338 }
3339
3340 if (error) {
3341 /*
3342 * Shutdown first so we kill the log before we release this
3343 * buffer. If it is an INODE_ALLOC buffer and pins the tail
3344 * of the log, failing it before the _log_ is shut down can
3345 * result in the log tail being moved forward in the journal
3346 * on disk because log writes can still be taking place. Hence
3347 * unpinning the tail will allow the ICREATE intent to be
3348 * removed from the log an recovery will fail with uninitialised
3349 * inode cluster buffers.
3350 */
3351 xfs_force_shutdown(mp, SHUTDOWN_CORRUPT_INCORE);
3352 bp->b_flags |= XBF_ASYNC;
3353 xfs_buf_ioend_fail(bp);
3354 return error;
3355 }
3356
3357 if (!clcount)
3358 return -EAGAIN;
3359
3360 XFS_STATS_INC(mp, xs_icluster_flushcnt);
3361 XFS_STATS_ADD(mp, xs_icluster_flushinode, clcount);
3362 return 0;
3363
3364 }
3365
3366 /* Release an inode. */
3367 void
xfs_irele(struct xfs_inode * ip)3368 xfs_irele(
3369 struct xfs_inode *ip)
3370 {
3371 trace_xfs_irele(ip, _RET_IP_);
3372 iput(VFS_I(ip));
3373 }
3374
3375 /*
3376 * Ensure all commited transactions touching the inode are written to the log.
3377 */
3378 int
xfs_log_force_inode(struct xfs_inode * ip)3379 xfs_log_force_inode(
3380 struct xfs_inode *ip)
3381 {
3382 xfs_csn_t seq = 0;
3383
3384 xfs_ilock(ip, XFS_ILOCK_SHARED);
3385 if (xfs_ipincount(ip))
3386 seq = ip->i_itemp->ili_commit_seq;
3387 xfs_iunlock(ip, XFS_ILOCK_SHARED);
3388
3389 if (!seq)
3390 return 0;
3391 return xfs_log_force_seq(ip->i_mount, seq, XFS_LOG_SYNC, NULL);
3392 }
3393
3394 /*
3395 * Grab the exclusive iolock for a data copy from src to dest, making sure to
3396 * abide vfs locking order (lowest pointer value goes first) and breaking the
3397 * layout leases before proceeding. The loop is needed because we cannot call
3398 * the blocking break_layout() with the iolocks held, and therefore have to
3399 * back out both locks.
3400 */
3401 static int
xfs_iolock_two_inodes_and_break_layout(struct inode * src,struct inode * dest)3402 xfs_iolock_two_inodes_and_break_layout(
3403 struct inode *src,
3404 struct inode *dest)
3405 {
3406 int error;
3407
3408 if (src > dest)
3409 swap(src, dest);
3410
3411 retry:
3412 /* Wait to break both inodes' layouts before we start locking. */
3413 error = break_layout(src, true);
3414 if (error)
3415 return error;
3416 if (src != dest) {
3417 error = break_layout(dest, true);
3418 if (error)
3419 return error;
3420 }
3421
3422 /* Lock one inode and make sure nobody got in and leased it. */
3423 inode_lock(src);
3424 error = break_layout(src, false);
3425 if (error) {
3426 inode_unlock(src);
3427 if (error == -EWOULDBLOCK)
3428 goto retry;
3429 return error;
3430 }
3431
3432 if (src == dest)
3433 return 0;
3434
3435 /* Lock the other inode and make sure nobody got in and leased it. */
3436 inode_lock_nested(dest, I_MUTEX_NONDIR2);
3437 error = break_layout(dest, false);
3438 if (error) {
3439 inode_unlock(src);
3440 inode_unlock(dest);
3441 if (error == -EWOULDBLOCK)
3442 goto retry;
3443 return error;
3444 }
3445
3446 return 0;
3447 }
3448
3449 static int
xfs_mmaplock_two_inodes_and_break_dax_layout(struct xfs_inode * ip1,struct xfs_inode * ip2)3450 xfs_mmaplock_two_inodes_and_break_dax_layout(
3451 struct xfs_inode *ip1,
3452 struct xfs_inode *ip2)
3453 {
3454 int error;
3455 bool retry;
3456 struct page *page;
3457
3458 if (ip1->i_ino > ip2->i_ino)
3459 swap(ip1, ip2);
3460
3461 again:
3462 retry = false;
3463 /* Lock the first inode */
3464 xfs_ilock(ip1, XFS_MMAPLOCK_EXCL);
3465 error = xfs_break_dax_layouts(VFS_I(ip1), &retry);
3466 if (error || retry) {
3467 xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
3468 if (error == 0 && retry)
3469 goto again;
3470 return error;
3471 }
3472
3473 if (ip1 == ip2)
3474 return 0;
3475
3476 /* Nested lock the second inode */
3477 xfs_ilock(ip2, xfs_lock_inumorder(XFS_MMAPLOCK_EXCL, 1));
3478 /*
3479 * We cannot use xfs_break_dax_layouts() directly here because it may
3480 * need to unlock & lock the XFS_MMAPLOCK_EXCL which is not suitable
3481 * for this nested lock case.
3482 */
3483 page = dax_layout_busy_page(VFS_I(ip2)->i_mapping);
3484 if (page && page_ref_count(page) != 1) {
3485 xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
3486 xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
3487 goto again;
3488 }
3489
3490 return 0;
3491 }
3492
3493 /*
3494 * Lock two inodes so that userspace cannot initiate I/O via file syscalls or
3495 * mmap activity.
3496 */
3497 int
xfs_ilock2_io_mmap(struct xfs_inode * ip1,struct xfs_inode * ip2)3498 xfs_ilock2_io_mmap(
3499 struct xfs_inode *ip1,
3500 struct xfs_inode *ip2)
3501 {
3502 int ret;
3503
3504 ret = xfs_iolock_two_inodes_and_break_layout(VFS_I(ip1), VFS_I(ip2));
3505 if (ret)
3506 return ret;
3507
3508 if (IS_DAX(VFS_I(ip1)) && IS_DAX(VFS_I(ip2))) {
3509 ret = xfs_mmaplock_two_inodes_and_break_dax_layout(ip1, ip2);
3510 if (ret) {
3511 inode_unlock(VFS_I(ip2));
3512 if (ip1 != ip2)
3513 inode_unlock(VFS_I(ip1));
3514 return ret;
3515 }
3516 } else
3517 filemap_invalidate_lock_two(VFS_I(ip1)->i_mapping,
3518 VFS_I(ip2)->i_mapping);
3519
3520 return 0;
3521 }
3522
3523 /* Unlock both inodes to allow IO and mmap activity. */
3524 void
xfs_iunlock2_io_mmap(struct xfs_inode * ip1,struct xfs_inode * ip2)3525 xfs_iunlock2_io_mmap(
3526 struct xfs_inode *ip1,
3527 struct xfs_inode *ip2)
3528 {
3529 if (IS_DAX(VFS_I(ip1)) && IS_DAX(VFS_I(ip2))) {
3530 xfs_iunlock(ip2, XFS_MMAPLOCK_EXCL);
3531 if (ip1 != ip2)
3532 xfs_iunlock(ip1, XFS_MMAPLOCK_EXCL);
3533 } else
3534 filemap_invalidate_unlock_two(VFS_I(ip1)->i_mapping,
3535 VFS_I(ip2)->i_mapping);
3536
3537 inode_unlock(VFS_I(ip2));
3538 if (ip1 != ip2)
3539 inode_unlock(VFS_I(ip1));
3540 }
3541