2 * Copyright (c) 2006-2007 Silicon Graphics, Inc.
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.
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
19 #include "xfs_bmap_btree.h"
21 #include "xfs_dinode.h"
22 #include "xfs_inode.h"
25 #include "xfs_trans.h"
27 #include "xfs_mount.h"
29 #include "xfs_alloc.h"
30 #include "xfs_utils.h"
31 #include "xfs_mru_cache.h"
32 #include "xfs_filestream.h"
33 #include "xfs_trace.h"
35 #ifdef XFS_FILESTREAMS_TRACE
37 ktrace_t *xfs_filestreams_trace_buf;
40 xfs_filestreams_trace(
41 xfs_mount_t *mp, /* mount point */
42 int type, /* type of trace */
43 const char *func, /* source function */
44 int line, /* source line number */
52 ktrace_enter(xfs_filestreams_trace_buf,
53 (void *)(__psint_t)(type | (line << 16)),
55 (void *)(__psunsigned_t)current_pid(),
57 (void *)(__psunsigned_t)arg0,
58 (void *)(__psunsigned_t)arg1,
59 (void *)(__psunsigned_t)arg2,
60 (void *)(__psunsigned_t)arg3,
61 (void *)(__psunsigned_t)arg4,
62 (void *)(__psunsigned_t)arg5,
63 NULL, NULL, NULL, NULL, NULL, NULL);
66 #define TRACE0(mp,t) TRACE6(mp,t,0,0,0,0,0,0)
67 #define TRACE1(mp,t,a0) TRACE6(mp,t,a0,0,0,0,0,0)
68 #define TRACE2(mp,t,a0,a1) TRACE6(mp,t,a0,a1,0,0,0,0)
69 #define TRACE3(mp,t,a0,a1,a2) TRACE6(mp,t,a0,a1,a2,0,0,0)
70 #define TRACE4(mp,t,a0,a1,a2,a3) TRACE6(mp,t,a0,a1,a2,a3,0,0)
71 #define TRACE5(mp,t,a0,a1,a2,a3,a4) TRACE6(mp,t,a0,a1,a2,a3,a4,0)
72 #define TRACE6(mp,t,a0,a1,a2,a3,a4,a5) \
73 xfs_filestreams_trace(mp, t, __func__, __LINE__, \
74 (__psunsigned_t)a0, (__psunsigned_t)a1, \
75 (__psunsigned_t)a2, (__psunsigned_t)a3, \
76 (__psunsigned_t)a4, (__psunsigned_t)a5)
78 #define TRACE_AG_SCAN(mp, ag, ag2) \
79 TRACE2(mp, XFS_FSTRM_KTRACE_AGSCAN, ag, ag2);
80 #define TRACE_AG_PICK1(mp, max_ag, maxfree) \
81 TRACE2(mp, XFS_FSTRM_KTRACE_AGPICK1, max_ag, maxfree);
82 #define TRACE_AG_PICK2(mp, ag, ag2, cnt, free, scan, flag) \
83 TRACE6(mp, XFS_FSTRM_KTRACE_AGPICK2, ag, ag2, \
84 cnt, free, scan, flag)
85 #define TRACE_UPDATE(mp, ip, ag, cnt, ag2, cnt2) \
86 TRACE5(mp, XFS_FSTRM_KTRACE_UPDATE, ip, ag, cnt, ag2, cnt2)
87 #define TRACE_FREE(mp, ip, pip, ag, cnt) \
88 TRACE4(mp, XFS_FSTRM_KTRACE_FREE, ip, pip, ag, cnt)
89 #define TRACE_LOOKUP(mp, ip, pip, ag, cnt) \
90 TRACE4(mp, XFS_FSTRM_KTRACE_ITEM_LOOKUP, ip, pip, ag, cnt)
91 #define TRACE_ASSOCIATE(mp, ip, pip, ag, cnt) \
92 TRACE4(mp, XFS_FSTRM_KTRACE_ASSOCIATE, ip, pip, ag, cnt)
93 #define TRACE_MOVEAG(mp, ip, pip, oag, ocnt, nag, ncnt) \
94 TRACE6(mp, XFS_FSTRM_KTRACE_MOVEAG, ip, pip, oag, ocnt, nag, ncnt)
95 #define TRACE_ORPHAN(mp, ip, ag) \
96 TRACE2(mp, XFS_FSTRM_KTRACE_ORPHAN, ip, ag);
100 #define TRACE_AG_SCAN(mp, ag, ag2)
101 #define TRACE_AG_PICK1(mp, max_ag, maxfree)
102 #define TRACE_AG_PICK2(mp, ag, ag2, cnt, free, scan, flag)
103 #define TRACE_UPDATE(mp, ip, ag, cnt, ag2, cnt2)
104 #define TRACE_FREE(mp, ip, pip, ag, cnt)
105 #define TRACE_LOOKUP(mp, ip, pip, ag, cnt)
106 #define TRACE_ASSOCIATE(mp, ip, pip, ag, cnt)
107 #define TRACE_MOVEAG(mp, ip, pip, oag, ocnt, nag, ncnt)
108 #define TRACE_ORPHAN(mp, ip, ag)
111 static kmem_zone_t *item_zone;
114 * Structure for associating a file or a directory with an allocation group.
115 * The parent directory pointer is only needed for files, but since there will
116 * generally be vastly more files than directories in the cache, using the same
117 * data structure simplifies the code with very little memory overhead.
119 typedef struct fstrm_item
121 xfs_agnumber_t ag; /* AG currently in use for the file/directory. */
122 xfs_inode_t *ip; /* inode self-pointer. */
123 xfs_inode_t *pip; /* Parent directory inode pointer. */
127 * Allocation group filestream associations are tracked with per-ag atomic
128 * counters. These counters allow _xfs_filestream_pick_ag() to tell whether a
129 * particular AG already has active filestreams associated with it. The mount
130 * point's m_peraglock is used to protect these counters from per-ag array
131 * re-allocation during a growfs operation. When xfs_growfs_data_private() is
132 * about to reallocate the array, it calls xfs_filestream_flush() with the
133 * m_peraglock held in write mode.
135 * Since xfs_mru_cache_flush() guarantees that all the free functions for all
136 * the cache elements have finished executing before it returns, it's safe for
137 * the free functions to use the atomic counters without m_peraglock protection.
138 * This allows the implementation of xfs_fstrm_free_func() to be agnostic about
139 * whether it was called with the m_peraglock held in read mode, write mode or
140 * not held at all. The race condition this addresses is the following:
142 * - The work queue scheduler fires and pulls a filestream directory cache
143 * element off the LRU end of the cache for deletion, then gets pre-empted.
144 * - A growfs operation grabs the m_peraglock in write mode, flushes all the
145 * remaining items from the cache and reallocates the mount point's per-ag
146 * array, resetting all the counters to zero.
147 * - The work queue thread resumes and calls the free function for the element
148 * it started cleaning up earlier. In the process it decrements the
149 * filestreams counter for an AG that now has no references.
151 * With a shrinkfs feature, the above scenario could panic the system.
153 * All other uses of the following macros should be protected by either the
154 * m_peraglock held in read mode, or the cache's internal locking exposed by the
155 * interval between a call to xfs_mru_cache_lookup() and a call to
156 * xfs_mru_cache_done(). In addition, the m_peraglock must be held in read mode
157 * when new elements are added to the cache.
159 * Combined, these locking rules ensure that no associations will ever exist in
160 * the cache that reference per-ag array elements that have since been
164 xfs_filestream_peek_ag(
168 struct xfs_perag *pag;
171 pag = xfs_perag_get(mp, agno);
172 ret = atomic_read(&pag->pagf_fstrms);
178 xfs_filestream_get_ag(
182 struct xfs_perag *pag;
185 pag = xfs_perag_get(mp, agno);
186 ret = atomic_inc_return(&pag->pagf_fstrms);
192 xfs_filestream_put_ag(
196 struct xfs_perag *pag;
198 pag = xfs_perag_get(mp, agno);
199 atomic_dec(&pag->pagf_fstrms);
204 * Scan the AGs starting at startag looking for an AG that isn't in use and has
205 * at least minlen blocks free.
208 _xfs_filestream_pick_ag(
210 xfs_agnumber_t startag,
215 int streams, max_streams;
216 int err, trylock, nscan;
217 xfs_extlen_t longest, free, minfree, maxfree = 0;
218 xfs_agnumber_t ag, max_ag = NULLAGNUMBER;
219 struct xfs_perag *pag;
221 /* 2% of an AG's blocks must be free for it to be chosen. */
222 minfree = mp->m_sb.sb_agblocks / 50;
227 /* For the first pass, don't sleep trying to init the per-AG. */
228 trylock = XFS_ALLOC_FLAG_TRYLOCK;
230 for (nscan = 0; 1; nscan++) {
231 pag = xfs_perag_get(mp, ag);
232 TRACE_AG_SCAN(mp, ag, atomic_read(&pag->pagf_fstrms));
234 if (!pag->pagf_init) {
235 err = xfs_alloc_pagf_init(mp, NULL, ag, trylock);
236 if (err && !trylock) {
242 /* Might fail sometimes during the 1st pass with trylock set. */
246 /* Keep track of the AG with the most free blocks. */
247 if (pag->pagf_freeblks > maxfree) {
248 maxfree = pag->pagf_freeblks;
249 max_streams = atomic_read(&pag->pagf_fstrms);
254 * The AG reference count does two things: it enforces mutual
255 * exclusion when examining the suitability of an AG in this
256 * loop, and it guards against two filestreams being established
257 * in the same AG as each other.
259 if (xfs_filestream_get_ag(mp, ag) > 1) {
260 xfs_filestream_put_ag(mp, ag);
264 longest = xfs_alloc_longest_free_extent(mp, pag);
265 if (((minlen && longest >= minlen) ||
266 (!minlen && pag->pagf_freeblks >= minfree)) &&
267 (!pag->pagf_metadata || !(flags & XFS_PICK_USERDATA) ||
268 (flags & XFS_PICK_LOWSPACE))) {
270 /* Break out, retaining the reference on the AG. */
271 free = pag->pagf_freeblks;
272 streams = atomic_read(&pag->pagf_fstrms);
278 /* Drop the reference on this AG, it's not usable. */
279 xfs_filestream_put_ag(mp, ag);
282 /* Move to the next AG, wrapping to AG 0 if necessary. */
283 if (++ag >= mp->m_sb.sb_agcount)
286 /* If a full pass of the AGs hasn't been done yet, continue. */
290 /* Allow sleeping in xfs_alloc_pagf_init() on the 2nd pass. */
296 /* Finally, if lowspace wasn't set, set it for the 3rd pass. */
297 if (!(flags & XFS_PICK_LOWSPACE)) {
298 flags |= XFS_PICK_LOWSPACE;
303 * Take the AG with the most free space, regardless of whether
304 * it's already in use by another filestream.
306 if (max_ag != NULLAGNUMBER) {
307 xfs_filestream_get_ag(mp, max_ag);
308 TRACE_AG_PICK1(mp, max_ag, maxfree);
309 streams = max_streams;
315 /* take AG 0 if none matched */
316 TRACE_AG_PICK1(mp, max_ag, maxfree);
321 TRACE_AG_PICK2(mp, startag, *agp, streams, free, nscan, flags);
327 * Set the allocation group number for a file or a directory, updating inode
328 * references and per-AG references as appropriate.
331 _xfs_filestream_update_ag(
338 xfs_mru_cache_t *cache;
340 xfs_agnumber_t old_ag;
341 xfs_inode_t *old_pip;
344 * Either ip is a regular file and pip is a directory, or ip is a
345 * directory and pip is NULL.
347 ASSERT(ip && ((S_ISREG(ip->i_d.di_mode) && pip &&
348 S_ISDIR(pip->i_d.di_mode)) ||
349 (S_ISDIR(ip->i_d.di_mode) && !pip)));
352 cache = mp->m_filestream;
354 item = xfs_mru_cache_lookup(cache, ip->i_ino);
356 ASSERT(item->ip == ip);
361 xfs_mru_cache_done(cache);
364 * If the AG has changed, drop the old ref and take a new one,
365 * effectively transferring the reference from old to new AG.
368 xfs_filestream_put_ag(mp, old_ag);
369 xfs_filestream_get_ag(mp, ag);
373 * If ip is a file and its pip has changed, drop the old ref and
376 if (pip && pip != old_pip) {
381 TRACE_UPDATE(mp, ip, old_ag, xfs_filestream_peek_ag(mp, old_ag),
382 ag, xfs_filestream_peek_ag(mp, ag));
386 item = kmem_zone_zalloc(item_zone, KM_MAYFAIL);
394 err = xfs_mru_cache_insert(cache, ip->i_ino, item);
396 kmem_zone_free(item_zone, item);
400 /* Take a reference on the AG. */
401 xfs_filestream_get_ag(mp, ag);
404 * Take a reference on the inode itself regardless of whether it's a
405 * regular file or a directory.
410 * In the case of a regular file, take a reference on the parent inode
411 * as well to ensure it remains in-core.
416 TRACE_UPDATE(mp, ip, ag, xfs_filestream_peek_ag(mp, ag),
417 ag, xfs_filestream_peek_ag(mp, ag));
422 /* xfs_fstrm_free_func(): callback for freeing cached stream items. */
428 fstrm_item_t *item = (fstrm_item_t *)data;
429 xfs_inode_t *ip = item->ip;
431 ASSERT(ip->i_ino == ino);
433 xfs_iflags_clear(ip, XFS_IFILESTREAM);
435 /* Drop the reference taken on the AG when the item was added. */
436 xfs_filestream_put_ag(ip->i_mount, item->ag);
438 TRACE_FREE(ip->i_mount, ip, item->pip, item->ag,
439 xfs_filestream_peek_ag(ip->i_mount, item->ag));
442 * _xfs_filestream_update_ag() always takes a reference on the inode
443 * itself, whether it's a file or a directory. Release it here.
444 * This can result in the inode being freed and so we must
445 * not hold any inode locks when freeing filesstreams objects
446 * otherwise we can deadlock here.
451 * In the case of a regular file, _xfs_filestream_update_ag() also
452 * takes a ref on the parent inode to keep it in-core. Release that
458 /* Finally, free the memory allocated for the item. */
459 kmem_zone_free(item_zone, item);
463 * xfs_filestream_init() is called at xfs initialisation time to set up the
464 * memory zone that will be used for filestream data structure allocation.
467 xfs_filestream_init(void)
469 item_zone = kmem_zone_init(sizeof(fstrm_item_t), "fstrm_item");
477 * xfs_filestream_uninit() is called at xfs termination time to destroy the
478 * memory zone that was used for filestream data structure allocation.
481 xfs_filestream_uninit(void)
483 kmem_zone_destroy(item_zone);
487 * xfs_filestream_mount() is called when a file system is mounted with the
488 * filestream option. It is responsible for allocating the data structures
489 * needed to track the new file system's file streams.
492 xfs_filestream_mount(
496 unsigned int lifetime, grp_count;
499 * The filestream timer tunable is currently fixed within the range of
500 * one second to four minutes, with five seconds being the default. The
501 * group count is somewhat arbitrary, but it'd be nice to adhere to the
502 * timer tunable to within about 10 percent. This requires at least 10
505 lifetime = xfs_fstrm_centisecs * 10;
508 err = xfs_mru_cache_create(&mp->m_filestream, lifetime, grp_count,
509 xfs_fstrm_free_func);
515 * xfs_filestream_unmount() is called when a file system that was mounted with
516 * the filestream option is unmounted. It drains the data structures created
517 * to track the file system's file streams and frees all the memory that was
521 xfs_filestream_unmount(
524 xfs_mru_cache_destroy(mp->m_filestream);
528 * Return the AG of the filestream the file or directory belongs to, or
529 * NULLAGNUMBER otherwise.
532 xfs_filestream_lookup_ag(
535 xfs_mru_cache_t *cache;
540 if (!S_ISREG(ip->i_d.di_mode) && !S_ISDIR(ip->i_d.di_mode)) {
545 cache = ip->i_mount->m_filestream;
546 item = xfs_mru_cache_lookup(cache, ip->i_ino);
548 TRACE_LOOKUP(ip->i_mount, ip, NULL, NULLAGNUMBER, 0);
552 ASSERT(ip == item->ip);
554 ref = xfs_filestream_peek_ag(ip->i_mount, ag);
555 xfs_mru_cache_done(cache);
557 TRACE_LOOKUP(ip->i_mount, ip, item->pip, ag, ref);
562 * xfs_filestream_associate() should only be called to associate a regular file
563 * with its parent directory. Calling it with a child directory isn't
564 * appropriate because filestreams don't apply to entire directory hierarchies.
565 * Creating a file in a child directory of an existing filestream directory
566 * starts a new filestream with its own allocation group association.
568 * Returns < 0 on error, 0 if successful association occurred, > 0 if
569 * we failed to get an association because of locking issues.
572 xfs_filestream_associate(
577 xfs_mru_cache_t *cache;
579 xfs_agnumber_t ag, rotorstep, startag;
582 ASSERT(S_ISDIR(pip->i_d.di_mode));
583 ASSERT(S_ISREG(ip->i_d.di_mode));
584 if (!S_ISDIR(pip->i_d.di_mode) || !S_ISREG(ip->i_d.di_mode))
588 cache = mp->m_filestream;
591 * We have a problem, Houston.
593 * Taking the iolock here violates inode locking order - we already
594 * hold the ilock. Hence if we block getting this lock we may never
595 * wake. Unfortunately, that means if we can't get the lock, we're
596 * screwed in terms of getting a stream association - we can't spin
597 * waiting for the lock because someone else is waiting on the lock we
598 * hold and we cannot drop that as we are in a transaction here.
600 * Lucky for us, this inversion is not a problem because it's a
601 * directory inode that we are trying to lock here.
603 * So, if we can't get the iolock without sleeping then just give up
605 if (!xfs_ilock_nowait(pip, XFS_IOLOCK_EXCL))
608 /* If the parent directory is already in the cache, use its AG. */
609 item = xfs_mru_cache_lookup(cache, pip->i_ino);
611 ASSERT(item->ip == pip);
613 xfs_mru_cache_done(cache);
615 TRACE_LOOKUP(mp, pip, pip, ag, xfs_filestream_peek_ag(mp, ag));
616 err = _xfs_filestream_update_ag(ip, pip, ag);
622 * Set the starting AG using the rotor for inode32, otherwise
623 * use the directory inode's AG.
625 if (mp->m_flags & XFS_MOUNT_32BITINODES) {
626 rotorstep = xfs_rotorstep;
627 startag = (mp->m_agfrotor / rotorstep) % mp->m_sb.sb_agcount;
628 mp->m_agfrotor = (mp->m_agfrotor + 1) %
629 (mp->m_sb.sb_agcount * rotorstep);
631 startag = XFS_INO_TO_AGNO(mp, pip->i_ino);
633 /* Pick a new AG for the parent inode starting at startag. */
634 err = _xfs_filestream_pick_ag(mp, startag, &ag, 0, 0);
635 if (err || ag == NULLAGNUMBER)
638 /* Associate the parent inode with the AG. */
639 err = _xfs_filestream_update_ag(pip, NULL, ag);
643 /* Associate the file inode with the AG. */
644 err = _xfs_filestream_update_ag(ip, pip, ag);
648 TRACE_ASSOCIATE(mp, ip, pip, ag, xfs_filestream_peek_ag(mp, ag));
652 * If _xfs_filestream_pick_ag() returned a valid AG, remove the
653 * reference it took on it, since the file and directory will have taken
654 * their own now if they were successfully cached.
656 if (ag != NULLAGNUMBER)
657 xfs_filestream_put_ag(mp, ag);
660 xfs_iunlock(pip, XFS_IOLOCK_EXCL);
665 * Pick a new allocation group for the current file and its file stream. This
666 * function is called by xfs_bmap_filestreams() with the mount point's per-ag
670 xfs_filestream_new_ag(
675 xfs_inode_t *ip, *pip = NULL;
677 xfs_mru_cache_t *cache;
679 fstrm_item_t *dir, *file;
680 xfs_agnumber_t ag = NULLAGNUMBER;
684 cache = mp->m_filestream;
689 * Look for the file in the cache, removing it if it's found. Doing
690 * this allows it to be held across the dir lookup that follows.
692 file = xfs_mru_cache_remove(cache, ip->i_ino);
694 ASSERT(ip == file->ip);
696 /* Save the file's parent inode and old AG number for later. */
700 /* Look for the file's directory in the cache. */
701 dir = xfs_mru_cache_lookup(cache, pip->i_ino);
703 ASSERT(pip == dir->ip);
706 * If the directory has already moved on to a new AG,
707 * use that AG as the new AG for the file. Don't
708 * forget to twiddle the AG refcounts to match the
711 if (dir->ag != file->ag) {
712 xfs_filestream_put_ag(mp, file->ag);
713 xfs_filestream_get_ag(mp, dir->ag);
714 *agp = file->ag = dir->ag;
717 xfs_mru_cache_done(cache);
721 * Put the file back in the cache. If this fails, the free
722 * function needs to be called to tidy up in the same way as if
723 * the item had simply expired from the cache.
725 err = xfs_mru_cache_insert(cache, ip->i_ino, file);
727 xfs_fstrm_free_func(ip->i_ino, file);
732 * If the file's AG was moved to the directory's new AG, there's
733 * nothing more to be done.
735 if (*agp != NULLAGNUMBER) {
736 TRACE_MOVEAG(mp, ip, pip,
737 ag, xfs_filestream_peek_ag(mp, ag),
738 *agp, xfs_filestream_peek_ag(mp, *agp));
744 * If the file's parent directory is known, take its iolock in exclusive
745 * mode to prevent two sibling files from racing each other to migrate
746 * themselves and their parent to different AGs.
748 * Note that we lock the parent directory iolock inside the child
749 * iolock here. That's fine as we never hold both parent and child
750 * iolock in any other place. This is different from the ilock,
751 * which requires locking of the child after the parent for namespace
755 xfs_ilock(pip, XFS_IOLOCK_EXCL | XFS_IOLOCK_PARENT);
758 * A new AG needs to be found for the file. If the file's parent
759 * directory is also known, it will be moved to the new AG as well to
760 * ensure that files created inside it in future use the new AG.
762 ag = (ag == NULLAGNUMBER) ? 0 : (ag + 1) % mp->m_sb.sb_agcount;
763 flags = (ap->userdata ? XFS_PICK_USERDATA : 0) |
764 (ap->flist->xbf_low ? XFS_PICK_LOWSPACE : 0);
766 err = _xfs_filestream_pick_ag(mp, ag, agp, flags, minlen);
767 if (err || *agp == NULLAGNUMBER)
771 * If the file wasn't found in the file cache, then its parent directory
772 * inode isn't known. For this to have happened, the file must either
773 * be pre-existing, or it was created long enough ago that its cache
774 * entry has expired. This isn't the sort of usage that the filestreams
775 * allocator is trying to optimise, so there's no point trying to track
776 * its new AG somehow in the filestream data structures.
779 TRACE_ORPHAN(mp, ip, *agp);
783 /* Associate the parent inode with the AG. */
784 err = _xfs_filestream_update_ag(pip, NULL, *agp);
788 /* Associate the file inode with the AG. */
789 err = _xfs_filestream_update_ag(ip, pip, *agp);
793 TRACE_MOVEAG(mp, ip, pip, NULLAGNUMBER, 0,
794 *agp, xfs_filestream_peek_ag(mp, *agp));
798 * If _xfs_filestream_pick_ag() returned a valid AG, remove the
799 * reference it took on it, since the file and directory will have taken
800 * their own now if they were successfully cached.
802 if (*agp != NULLAGNUMBER)
803 xfs_filestream_put_ag(mp, *agp);
808 xfs_iunlock(pip, XFS_IOLOCK_EXCL);
814 * Remove an association between an inode and a filestream object.
815 * Typically this is done on last close of an unlinked file.
818 xfs_filestream_deassociate(
821 xfs_mru_cache_t *cache = ip->i_mount->m_filestream;
823 xfs_mru_cache_delete(cache, ip->i_ino);