xfs_icache.c (6d8b79cfca39399ef9115fb65dde85993455c9a3) xfs_icache.c (33479e0542df066fb0b47df18780e93bfe6e0dc5)
1/*
2 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
3 * All Rights Reserved.
4 *
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
8 *

--- 27 unchanged lines hidden (view full) ---

36#include "xfs_quota.h"
37#include "xfs_trace.h"
38#include "xfs_fsops.h"
39#include "xfs_icache.h"
40
41#include <linux/kthread.h>
42#include <linux/freezer.h>
43
1/*
2 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
3 * All Rights Reserved.
4 *
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
8 *

--- 27 unchanged lines hidden (view full) ---

36#include "xfs_quota.h"
37#include "xfs_trace.h"
38#include "xfs_fsops.h"
39#include "xfs_icache.h"
40
41#include <linux/kthread.h>
42#include <linux/freezer.h>
43
44STATIC void __xfs_inode_clear_reclaim_tag(struct xfs_mount *mp,
45 struct xfs_perag *pag, struct xfs_inode *ip);
46
44/*
47/*
48 * Allocate and initialise an xfs_inode.
49 */
50STATIC struct xfs_inode *
51xfs_inode_alloc(
52 struct xfs_mount *mp,
53 xfs_ino_t ino)
54{
55 struct xfs_inode *ip;
56
57 /*
58 * if this didn't occur in transactions, we could use
59 * KM_MAYFAIL and return NULL here on ENOMEM. Set the
60 * code up to do this anyway.
61 */
62 ip = kmem_zone_alloc(xfs_inode_zone, KM_SLEEP);
63 if (!ip)
64 return NULL;
65 if (inode_init_always(mp->m_super, VFS_I(ip))) {
66 kmem_zone_free(xfs_inode_zone, ip);
67 return NULL;
68 }
69
70 ASSERT(atomic_read(&ip->i_pincount) == 0);
71 ASSERT(!spin_is_locked(&ip->i_flags_lock));
72 ASSERT(!xfs_isiflocked(ip));
73 ASSERT(ip->i_ino == 0);
74
75 mrlock_init(&ip->i_iolock, MRLOCK_BARRIER, "xfsio", ip->i_ino);
76
77 /* initialise the xfs inode */
78 ip->i_ino = ino;
79 ip->i_mount = mp;
80 memset(&ip->i_imap, 0, sizeof(struct xfs_imap));
81 ip->i_afp = NULL;
82 memset(&ip->i_df, 0, sizeof(xfs_ifork_t));
83 ip->i_flags = 0;
84 ip->i_delayed_blks = 0;
85 memset(&ip->i_d, 0, sizeof(xfs_icdinode_t));
86
87 return ip;
88}
89
90STATIC void
91xfs_inode_free_callback(
92 struct rcu_head *head)
93{
94 struct inode *inode = container_of(head, struct inode, i_rcu);
95 struct xfs_inode *ip = XFS_I(inode);
96
97 kmem_zone_free(xfs_inode_zone, ip);
98}
99
100STATIC void
101xfs_inode_free(
102 struct xfs_inode *ip)
103{
104 switch (ip->i_d.di_mode & S_IFMT) {
105 case S_IFREG:
106 case S_IFDIR:
107 case S_IFLNK:
108 xfs_idestroy_fork(ip, XFS_DATA_FORK);
109 break;
110 }
111
112 if (ip->i_afp)
113 xfs_idestroy_fork(ip, XFS_ATTR_FORK);
114
115 if (ip->i_itemp) {
116 ASSERT(!(ip->i_itemp->ili_item.li_flags & XFS_LI_IN_AIL));
117 xfs_inode_item_destroy(ip);
118 ip->i_itemp = NULL;
119 }
120
121 /* asserts to verify all state is correct here */
122 ASSERT(atomic_read(&ip->i_pincount) == 0);
123 ASSERT(!spin_is_locked(&ip->i_flags_lock));
124 ASSERT(!xfs_isiflocked(ip));
125
126 /*
127 * Because we use RCU freeing we need to ensure the inode always
128 * appears to be reclaimed with an invalid inode number when in the
129 * free state. The ip->i_flags_lock provides the barrier against lookup
130 * races.
131 */
132 spin_lock(&ip->i_flags_lock);
133 ip->i_flags = XFS_IRECLAIM;
134 ip->i_ino = 0;
135 spin_unlock(&ip->i_flags_lock);
136
137 call_rcu(&VFS_I(ip)->i_rcu, xfs_inode_free_callback);
138}
139
140/*
141 * Check the validity of the inode we just found it the cache
142 */
143static int
144xfs_iget_cache_hit(
145 struct xfs_perag *pag,
146 struct xfs_inode *ip,
147 xfs_ino_t ino,
148 int flags,
149 int lock_flags) __releases(RCU)
150{
151 struct inode *inode = VFS_I(ip);
152 struct xfs_mount *mp = ip->i_mount;
153 int error;
154
155 /*
156 * check for re-use of an inode within an RCU grace period due to the
157 * radix tree nodes not being updated yet. We monitor for this by
158 * setting the inode number to zero before freeing the inode structure.
159 * If the inode has been reallocated and set up, then the inode number
160 * will not match, so check for that, too.
161 */
162 spin_lock(&ip->i_flags_lock);
163 if (ip->i_ino != ino) {
164 trace_xfs_iget_skip(ip);
165 XFS_STATS_INC(xs_ig_frecycle);
166 error = EAGAIN;
167 goto out_error;
168 }
169
170
171 /*
172 * If we are racing with another cache hit that is currently
173 * instantiating this inode or currently recycling it out of
174 * reclaimabe state, wait for the initialisation to complete
175 * before continuing.
176 *
177 * XXX(hch): eventually we should do something equivalent to
178 * wait_on_inode to wait for these flags to be cleared
179 * instead of polling for it.
180 */
181 if (ip->i_flags & (XFS_INEW|XFS_IRECLAIM)) {
182 trace_xfs_iget_skip(ip);
183 XFS_STATS_INC(xs_ig_frecycle);
184 error = EAGAIN;
185 goto out_error;
186 }
187
188 /*
189 * If lookup is racing with unlink return an error immediately.
190 */
191 if (ip->i_d.di_mode == 0 && !(flags & XFS_IGET_CREATE)) {
192 error = ENOENT;
193 goto out_error;
194 }
195
196 /*
197 * If IRECLAIMABLE is set, we've torn down the VFS inode already.
198 * Need to carefully get it back into useable state.
199 */
200 if (ip->i_flags & XFS_IRECLAIMABLE) {
201 trace_xfs_iget_reclaim(ip);
202
203 /*
204 * We need to set XFS_IRECLAIM to prevent xfs_reclaim_inode
205 * from stomping over us while we recycle the inode. We can't
206 * clear the radix tree reclaimable tag yet as it requires
207 * pag_ici_lock to be held exclusive.
208 */
209 ip->i_flags |= XFS_IRECLAIM;
210
211 spin_unlock(&ip->i_flags_lock);
212 rcu_read_unlock();
213
214 error = -inode_init_always(mp->m_super, inode);
215 if (error) {
216 /*
217 * Re-initializing the inode failed, and we are in deep
218 * trouble. Try to re-add it to the reclaim list.
219 */
220 rcu_read_lock();
221 spin_lock(&ip->i_flags_lock);
222
223 ip->i_flags &= ~(XFS_INEW | XFS_IRECLAIM);
224 ASSERT(ip->i_flags & XFS_IRECLAIMABLE);
225 trace_xfs_iget_reclaim_fail(ip);
226 goto out_error;
227 }
228
229 spin_lock(&pag->pag_ici_lock);
230 spin_lock(&ip->i_flags_lock);
231
232 /*
233 * Clear the per-lifetime state in the inode as we are now
234 * effectively a new inode and need to return to the initial
235 * state before reuse occurs.
236 */
237 ip->i_flags &= ~XFS_IRECLAIM_RESET_FLAGS;
238 ip->i_flags |= XFS_INEW;
239 __xfs_inode_clear_reclaim_tag(mp, pag, ip);
240 inode->i_state = I_NEW;
241
242 ASSERT(!rwsem_is_locked(&ip->i_iolock.mr_lock));
243 mrlock_init(&ip->i_iolock, MRLOCK_BARRIER, "xfsio", ip->i_ino);
244
245 spin_unlock(&ip->i_flags_lock);
246 spin_unlock(&pag->pag_ici_lock);
247 } else {
248 /* If the VFS inode is being torn down, pause and try again. */
249 if (!igrab(inode)) {
250 trace_xfs_iget_skip(ip);
251 error = EAGAIN;
252 goto out_error;
253 }
254
255 /* We've got a live one. */
256 spin_unlock(&ip->i_flags_lock);
257 rcu_read_unlock();
258 trace_xfs_iget_hit(ip);
259 }
260
261 if (lock_flags != 0)
262 xfs_ilock(ip, lock_flags);
263
264 xfs_iflags_clear(ip, XFS_ISTALE | XFS_IDONTCACHE);
265 XFS_STATS_INC(xs_ig_found);
266
267 return 0;
268
269out_error:
270 spin_unlock(&ip->i_flags_lock);
271 rcu_read_unlock();
272 return error;
273}
274
275
276static int
277xfs_iget_cache_miss(
278 struct xfs_mount *mp,
279 struct xfs_perag *pag,
280 xfs_trans_t *tp,
281 xfs_ino_t ino,
282 struct xfs_inode **ipp,
283 int flags,
284 int lock_flags)
285{
286 struct xfs_inode *ip;
287 int error;
288 xfs_agino_t agino = XFS_INO_TO_AGINO(mp, ino);
289 int iflags;
290
291 ip = xfs_inode_alloc(mp, ino);
292 if (!ip)
293 return ENOMEM;
294
295 error = xfs_iread(mp, tp, ip, flags);
296 if (error)
297 goto out_destroy;
298
299 trace_xfs_iget_miss(ip);
300
301 if ((ip->i_d.di_mode == 0) && !(flags & XFS_IGET_CREATE)) {
302 error = ENOENT;
303 goto out_destroy;
304 }
305
306 /*
307 * Preload the radix tree so we can insert safely under the
308 * write spinlock. Note that we cannot sleep inside the preload
309 * region. Since we can be called from transaction context, don't
310 * recurse into the file system.
311 */
312 if (radix_tree_preload(GFP_NOFS)) {
313 error = EAGAIN;
314 goto out_destroy;
315 }
316
317 /*
318 * Because the inode hasn't been added to the radix-tree yet it can't
319 * be found by another thread, so we can do the non-sleeping lock here.
320 */
321 if (lock_flags) {
322 if (!xfs_ilock_nowait(ip, lock_flags))
323 BUG();
324 }
325
326 /*
327 * These values must be set before inserting the inode into the radix
328 * tree as the moment it is inserted a concurrent lookup (allowed by the
329 * RCU locking mechanism) can find it and that lookup must see that this
330 * is an inode currently under construction (i.e. that XFS_INEW is set).
331 * The ip->i_flags_lock that protects the XFS_INEW flag forms the
332 * memory barrier that ensures this detection works correctly at lookup
333 * time.
334 */
335 iflags = XFS_INEW;
336 if (flags & XFS_IGET_DONTCACHE)
337 iflags |= XFS_IDONTCACHE;
338 ip->i_udquot = ip->i_gdquot = NULL;
339 xfs_iflags_set(ip, iflags);
340
341 /* insert the new inode */
342 spin_lock(&pag->pag_ici_lock);
343 error = radix_tree_insert(&pag->pag_ici_root, agino, ip);
344 if (unlikely(error)) {
345 WARN_ON(error != -EEXIST);
346 XFS_STATS_INC(xs_ig_dup);
347 error = EAGAIN;
348 goto out_preload_end;
349 }
350 spin_unlock(&pag->pag_ici_lock);
351 radix_tree_preload_end();
352
353 *ipp = ip;
354 return 0;
355
356out_preload_end:
357 spin_unlock(&pag->pag_ici_lock);
358 radix_tree_preload_end();
359 if (lock_flags)
360 xfs_iunlock(ip, lock_flags);
361out_destroy:
362 __destroy_inode(VFS_I(ip));
363 xfs_inode_free(ip);
364 return error;
365}
366
367/*
368 * Look up an inode by number in the given file system.
369 * The inode is looked up in the cache held in each AG.
370 * If the inode is found in the cache, initialise the vfs inode
371 * if necessary.
372 *
373 * If it is not in core, read it in from the file system's device,
374 * add it to the cache and initialise the vfs inode.
375 *
376 * The inode is locked according to the value of the lock_flags parameter.
377 * This flag parameter indicates how and if the inode's IO lock and inode lock
378 * should be taken.
379 *
380 * mp -- the mount point structure for the current file system. It points
381 * to the inode hash table.
382 * tp -- a pointer to the current transaction if there is one. This is
383 * simply passed through to the xfs_iread() call.
384 * ino -- the number of the inode desired. This is the unique identifier
385 * within the file system for the inode being requested.
386 * lock_flags -- flags indicating how to lock the inode. See the comment
387 * for xfs_ilock() for a list of valid values.
388 */
389int
390xfs_iget(
391 xfs_mount_t *mp,
392 xfs_trans_t *tp,
393 xfs_ino_t ino,
394 uint flags,
395 uint lock_flags,
396 xfs_inode_t **ipp)
397{
398 xfs_inode_t *ip;
399 int error;
400 xfs_perag_t *pag;
401 xfs_agino_t agino;
402
403 /*
404 * xfs_reclaim_inode() uses the ILOCK to ensure an inode
405 * doesn't get freed while it's being referenced during a
406 * radix tree traversal here. It assumes this function
407 * aqcuires only the ILOCK (and therefore it has no need to
408 * involve the IOLOCK in this synchronization).
409 */
410 ASSERT((lock_flags & (XFS_IOLOCK_EXCL | XFS_IOLOCK_SHARED)) == 0);
411
412 /* reject inode numbers outside existing AGs */
413 if (!ino || XFS_INO_TO_AGNO(mp, ino) >= mp->m_sb.sb_agcount)
414 return EINVAL;
415
416 /* get the perag structure and ensure that it's inode capable */
417 pag = xfs_perag_get(mp, XFS_INO_TO_AGNO(mp, ino));
418 agino = XFS_INO_TO_AGINO(mp, ino);
419
420again:
421 error = 0;
422 rcu_read_lock();
423 ip = radix_tree_lookup(&pag->pag_ici_root, agino);
424
425 if (ip) {
426 error = xfs_iget_cache_hit(pag, ip, ino, flags, lock_flags);
427 if (error)
428 goto out_error_or_again;
429 } else {
430 rcu_read_unlock();
431 XFS_STATS_INC(xs_ig_missed);
432
433 error = xfs_iget_cache_miss(mp, pag, tp, ino, &ip,
434 flags, lock_flags);
435 if (error)
436 goto out_error_or_again;
437 }
438 xfs_perag_put(pag);
439
440 *ipp = ip;
441
442 /*
443 * If we have a real type for an on-disk inode, we can set ops(&unlock)
444 * now. If it's a new inode being created, xfs_ialloc will handle it.
445 */
446 if (xfs_iflags_test(ip, XFS_INEW) && ip->i_d.di_mode != 0)
447 xfs_setup_inode(ip);
448 return 0;
449
450out_error_or_again:
451 if (error == EAGAIN) {
452 delay(1);
453 goto again;
454 }
455 xfs_perag_put(pag);
456 return error;
457}
458
459/*
45 * The inode lookup is done in batches to keep the amount of lock traffic and
46 * radix tree lookups to a minimum. The batch size is a trade off between
47 * lookup reduction and stack usage. This is in the reclaim path, so we can't
48 * be too greedy.
49 */
50#define XFS_LOOKUP_BATCH 32
51
52STATIC int

--- 195 unchanged lines hidden (view full) ---

248{
249 struct xfs_mount *mp = container_of(to_delayed_work(work),
250 struct xfs_mount, m_reclaim_work);
251
252 xfs_reclaim_inodes(mp, SYNC_TRYLOCK);
253 xfs_reclaim_work_queue(mp);
254}
255
460 * The inode lookup is done in batches to keep the amount of lock traffic and
461 * radix tree lookups to a minimum. The batch size is a trade off between
462 * lookup reduction and stack usage. This is in the reclaim path, so we can't
463 * be too greedy.
464 */
465#define XFS_LOOKUP_BATCH 32
466
467STATIC int

--- 195 unchanged lines hidden (view full) ---

663{
664 struct xfs_mount *mp = container_of(to_delayed_work(work),
665 struct xfs_mount, m_reclaim_work);
666
667 xfs_reclaim_inodes(mp, SYNC_TRYLOCK);
668 xfs_reclaim_work_queue(mp);
669}
670
256void
671static void
257__xfs_inode_set_reclaim_tag(
258 struct xfs_perag *pag,
259 struct xfs_inode *ip)
260{
261 radix_tree_tag_set(&pag->pag_ici_root,
262 XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino),
263 XFS_ICI_RECLAIM_TAG);
264

--- 49 unchanged lines hidden (view full) ---

314 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
315 XFS_ICI_RECLAIM_TAG);
316 spin_unlock(&ip->i_mount->m_perag_lock);
317 trace_xfs_perag_clear_reclaim(ip->i_mount, pag->pag_agno,
318 -1, _RET_IP_);
319 }
320}
321
672__xfs_inode_set_reclaim_tag(
673 struct xfs_perag *pag,
674 struct xfs_inode *ip)
675{
676 radix_tree_tag_set(&pag->pag_ici_root,
677 XFS_INO_TO_AGINO(ip->i_mount, ip->i_ino),
678 XFS_ICI_RECLAIM_TAG);
679

--- 49 unchanged lines hidden (view full) ---

729 XFS_INO_TO_AGNO(ip->i_mount, ip->i_ino),
730 XFS_ICI_RECLAIM_TAG);
731 spin_unlock(&ip->i_mount->m_perag_lock);
732 trace_xfs_perag_clear_reclaim(ip->i_mount, pag->pag_agno,
733 -1, _RET_IP_);
734 }
735}
736
322void
737STATIC void
323__xfs_inode_clear_reclaim_tag(
324 xfs_mount_t *mp,
325 xfs_perag_t *pag,
326 xfs_inode_t *ip)
327{
328 radix_tree_tag_clear(&pag->pag_ici_root,
329 XFS_INO_TO_AGINO(mp, ip->i_ino), XFS_ICI_RECLAIM_TAG);
330 __xfs_inode_clear_reclaim(pag, ip);

--- 206 unchanged lines hidden (view full) ---

537}
538
539/*
540 * Walk the AGs and reclaim the inodes in them. Even if the filesystem is
541 * corrupted, we still want to try to reclaim all the inodes. If we don't,
542 * then a shut down during filesystem unmount reclaim walk leak all the
543 * unreclaimed inodes.
544 */
738__xfs_inode_clear_reclaim_tag(
739 xfs_mount_t *mp,
740 xfs_perag_t *pag,
741 xfs_inode_t *ip)
742{
743 radix_tree_tag_clear(&pag->pag_ici_root,
744 XFS_INO_TO_AGINO(mp, ip->i_ino), XFS_ICI_RECLAIM_TAG);
745 __xfs_inode_clear_reclaim(pag, ip);

--- 206 unchanged lines hidden (view full) ---

952}
953
954/*
955 * Walk the AGs and reclaim the inodes in them. Even if the filesystem is
956 * corrupted, we still want to try to reclaim all the inodes. If we don't,
957 * then a shut down during filesystem unmount reclaim walk leak all the
958 * unreclaimed inodes.
959 */
545int
960STATIC int
546xfs_reclaim_inodes_ag(
547 struct xfs_mount *mp,
548 int flags,
549 int *nr_to_scan)
550{
551 struct xfs_perag *pag;
552 int error = 0;
553 int last_error = 0;

--- 162 unchanged lines hidden --- 		961xfs_reclaim_inodes_ag(
962 struct xfs_mount *mp,
963 int flags,
964 int *nr_to_scan)
965{
966 struct xfs_perag *pag;
967 int error = 0;
968 int last_error = 0;

--- 162 unchanged lines hidden ---