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"
22 #include "xfs_dir2_sf.h"
23 #include "xfs_attr_sf.h"
24 #include "xfs_dinode.h"
25 #include "xfs_inode.h"
27 #include "xfs_dmapi.h"
29 #include "xfs_trans.h"
31 #include "xfs_mount.h"
33 #include "xfs_alloc.h"
34 #include "xfs_utils.h"
35 #include "xfs_mru_cache.h"
36 #include "xfs_filestream.h"
37 #include "xfs_trace.h"
39 #ifdef XFS_FILESTREAMS_TRACE
41 ktrace_t
*xfs_filestreams_trace_buf
;
44 xfs_filestreams_trace(
45 xfs_mount_t
*mp
, /* mount point */
46 int type
, /* type of trace */
47 const char *func
, /* source function */
48 int line
, /* source line number */
56 ktrace_enter(xfs_filestreams_trace_buf
,
57 (void *)(__psint_t
)(type
| (line
<< 16)),
59 (void *)(__psunsigned_t
)current_pid(),
61 (void *)(__psunsigned_t
)arg0
,
62 (void *)(__psunsigned_t
)arg1
,
63 (void *)(__psunsigned_t
)arg2
,
64 (void *)(__psunsigned_t
)arg3
,
65 (void *)(__psunsigned_t
)arg4
,
66 (void *)(__psunsigned_t
)arg5
,
67 NULL
, NULL
, NULL
, NULL
, NULL
, NULL
);
70 #define TRACE0(mp,t) TRACE6(mp,t,0,0,0,0,0,0)
71 #define TRACE1(mp,t,a0) TRACE6(mp,t,a0,0,0,0,0,0)
72 #define TRACE2(mp,t,a0,a1) TRACE6(mp,t,a0,a1,0,0,0,0)
73 #define TRACE3(mp,t,a0,a1,a2) TRACE6(mp,t,a0,a1,a2,0,0,0)
74 #define TRACE4(mp,t,a0,a1,a2,a3) TRACE6(mp,t,a0,a1,a2,a3,0,0)
75 #define TRACE5(mp,t,a0,a1,a2,a3,a4) TRACE6(mp,t,a0,a1,a2,a3,a4,0)
76 #define TRACE6(mp,t,a0,a1,a2,a3,a4,a5) \
77 xfs_filestreams_trace(mp, t, __func__, __LINE__, \
78 (__psunsigned_t)a0, (__psunsigned_t)a1, \
79 (__psunsigned_t)a2, (__psunsigned_t)a3, \
80 (__psunsigned_t)a4, (__psunsigned_t)a5)
82 #define TRACE_AG_SCAN(mp, ag, ag2) \
83 TRACE2(mp, XFS_FSTRM_KTRACE_AGSCAN, ag, ag2);
84 #define TRACE_AG_PICK1(mp, max_ag, maxfree) \
85 TRACE2(mp, XFS_FSTRM_KTRACE_AGPICK1, max_ag, maxfree);
86 #define TRACE_AG_PICK2(mp, ag, ag2, cnt, free, scan, flag) \
87 TRACE6(mp, XFS_FSTRM_KTRACE_AGPICK2, ag, ag2, \
88 cnt, free, scan, flag)
89 #define TRACE_UPDATE(mp, ip, ag, cnt, ag2, cnt2) \
90 TRACE5(mp, XFS_FSTRM_KTRACE_UPDATE, ip, ag, cnt, ag2, cnt2)
91 #define TRACE_FREE(mp, ip, pip, ag, cnt) \
92 TRACE4(mp, XFS_FSTRM_KTRACE_FREE, ip, pip, ag, cnt)
93 #define TRACE_LOOKUP(mp, ip, pip, ag, cnt) \
94 TRACE4(mp, XFS_FSTRM_KTRACE_ITEM_LOOKUP, ip, pip, ag, cnt)
95 #define TRACE_ASSOCIATE(mp, ip, pip, ag, cnt) \
96 TRACE4(mp, XFS_FSTRM_KTRACE_ASSOCIATE, ip, pip, ag, cnt)
97 #define TRACE_MOVEAG(mp, ip, pip, oag, ocnt, nag, ncnt) \
98 TRACE6(mp, XFS_FSTRM_KTRACE_MOVEAG, ip, pip, oag, ocnt, nag, ncnt)
99 #define TRACE_ORPHAN(mp, ip, ag) \
100 TRACE2(mp, XFS_FSTRM_KTRACE_ORPHAN, ip, ag);
104 #define TRACE_AG_SCAN(mp, ag, ag2)
105 #define TRACE_AG_PICK1(mp, max_ag, maxfree)
106 #define TRACE_AG_PICK2(mp, ag, ag2, cnt, free, scan, flag)
107 #define TRACE_UPDATE(mp, ip, ag, cnt, ag2, cnt2)
108 #define TRACE_FREE(mp, ip, pip, ag, cnt)
109 #define TRACE_LOOKUP(mp, ip, pip, ag, cnt)
110 #define TRACE_ASSOCIATE(mp, ip, pip, ag, cnt)
111 #define TRACE_MOVEAG(mp, ip, pip, oag, ocnt, nag, ncnt)
112 #define TRACE_ORPHAN(mp, ip, ag)
115 static kmem_zone_t
*item_zone
;
118 * Structure for associating a file or a directory with an allocation group.
119 * The parent directory pointer is only needed for files, but since there will
120 * generally be vastly more files than directories in the cache, using the same
121 * data structure simplifies the code with very little memory overhead.
123 typedef struct fstrm_item
125 xfs_agnumber_t ag
; /* AG currently in use for the file/directory. */
126 xfs_inode_t
*ip
; /* inode self-pointer. */
127 xfs_inode_t
*pip
; /* Parent directory inode pointer. */
132 * Scan the AGs starting at startag looking for an AG that isn't in use and has
133 * at least minlen blocks free.
136 _xfs_filestream_pick_ag(
138 xfs_agnumber_t startag
,
143 int streams
, max_streams
;
144 int err
, trylock
, nscan
;
145 xfs_extlen_t longest
, free
, minfree
, maxfree
= 0;
146 xfs_agnumber_t ag
, max_ag
= NULLAGNUMBER
;
147 struct xfs_perag
*pag
;
149 /* 2% of an AG's blocks must be free for it to be chosen. */
150 minfree
= mp
->m_sb
.sb_agblocks
/ 50;
155 /* For the first pass, don't sleep trying to init the per-AG. */
156 trylock
= XFS_ALLOC_FLAG_TRYLOCK
;
158 for (nscan
= 0; 1; nscan
++) {
159 pag
= xfs_perag_get(mp
, ag
);
160 TRACE_AG_SCAN(mp
, ag
, atomic_read(&pag
->pagf_fstrms
));
162 if (!pag
->pagf_init
) {
163 err
= xfs_alloc_pagf_init(mp
, NULL
, ag
, trylock
);
164 if (err
&& !trylock
) {
170 /* Might fail sometimes during the 1st pass with trylock set. */
174 /* Keep track of the AG with the most free blocks. */
175 if (pag
->pagf_freeblks
> maxfree
) {
176 maxfree
= pag
->pagf_freeblks
;
177 max_streams
= atomic_read(&pag
->pagf_fstrms
);
182 * The AG reference count does two things: it enforces mutual
183 * exclusion when examining the suitability of an AG in this
184 * loop, and it guards against two filestreams being established
185 * in the same AG as each other.
187 if (xfs_filestream_get_ag(mp
, ag
) > 1) {
188 xfs_filestream_put_ag(mp
, ag
);
192 longest
= xfs_alloc_longest_free_extent(mp
, pag
);
193 if (((minlen
&& longest
>= minlen
) ||
194 (!minlen
&& pag
->pagf_freeblks
>= minfree
)) &&
195 (!pag
->pagf_metadata
|| !(flags
& XFS_PICK_USERDATA
) ||
196 (flags
& XFS_PICK_LOWSPACE
))) {
198 /* Break out, retaining the reference on the AG. */
199 free
= pag
->pagf_freeblks
;
200 streams
= atomic_read(&pag
->pagf_fstrms
);
206 /* Drop the reference on this AG, it's not usable. */
207 xfs_filestream_put_ag(mp
, ag
);
210 /* Move to the next AG, wrapping to AG 0 if necessary. */
211 if (++ag
>= mp
->m_sb
.sb_agcount
)
214 /* If a full pass of the AGs hasn't been done yet, continue. */
218 /* Allow sleeping in xfs_alloc_pagf_init() on the 2nd pass. */
224 /* Finally, if lowspace wasn't set, set it for the 3rd pass. */
225 if (!(flags
& XFS_PICK_LOWSPACE
)) {
226 flags
|= XFS_PICK_LOWSPACE
;
231 * Take the AG with the most free space, regardless of whether
232 * it's already in use by another filestream.
234 if (max_ag
!= NULLAGNUMBER
) {
235 xfs_filestream_get_ag(mp
, max_ag
);
236 TRACE_AG_PICK1(mp
, max_ag
, maxfree
);
237 streams
= max_streams
;
243 /* take AG 0 if none matched */
244 TRACE_AG_PICK1(mp
, max_ag
, maxfree
);
249 TRACE_AG_PICK2(mp
, startag
, *agp
, streams
, free
, nscan
, flags
);
255 * Set the allocation group number for a file or a directory, updating inode
256 * references and per-AG references as appropriate.
259 _xfs_filestream_update_ag(
266 xfs_mru_cache_t
*cache
;
268 xfs_agnumber_t old_ag
;
269 xfs_inode_t
*old_pip
;
272 * Either ip is a regular file and pip is a directory, or ip is a
273 * directory and pip is NULL.
275 ASSERT(ip
&& (((ip
->i_d
.di_mode
& S_IFREG
) && pip
&&
276 (pip
->i_d
.di_mode
& S_IFDIR
)) ||
277 ((ip
->i_d
.di_mode
& S_IFDIR
) && !pip
)));
280 cache
= mp
->m_filestream
;
282 item
= xfs_mru_cache_lookup(cache
, ip
->i_ino
);
284 ASSERT(item
->ip
== ip
);
289 xfs_mru_cache_done(cache
);
292 * If the AG has changed, drop the old ref and take a new one,
293 * effectively transferring the reference from old to new AG.
296 xfs_filestream_put_ag(mp
, old_ag
);
297 xfs_filestream_get_ag(mp
, ag
);
301 * If ip is a file and its pip has changed, drop the old ref and
304 if (pip
&& pip
!= old_pip
) {
309 TRACE_UPDATE(mp
, ip
, old_ag
, xfs_filestream_peek_ag(mp
, old_ag
),
310 ag
, xfs_filestream_peek_ag(mp
, ag
));
314 item
= kmem_zone_zalloc(item_zone
, KM_MAYFAIL
);
322 err
= xfs_mru_cache_insert(cache
, ip
->i_ino
, item
);
324 kmem_zone_free(item_zone
, item
);
328 /* Take a reference on the AG. */
329 xfs_filestream_get_ag(mp
, ag
);
332 * Take a reference on the inode itself regardless of whether it's a
333 * regular file or a directory.
338 * In the case of a regular file, take a reference on the parent inode
339 * as well to ensure it remains in-core.
344 TRACE_UPDATE(mp
, ip
, ag
, xfs_filestream_peek_ag(mp
, ag
),
345 ag
, xfs_filestream_peek_ag(mp
, ag
));
350 /* xfs_fstrm_free_func(): callback for freeing cached stream items. */
356 fstrm_item_t
*item
= (fstrm_item_t
*)data
;
357 xfs_inode_t
*ip
= item
->ip
;
360 ASSERT(ip
->i_ino
== ino
);
362 xfs_iflags_clear(ip
, XFS_IFILESTREAM
);
364 /* Drop the reference taken on the AG when the item was added. */
365 ref
= xfs_filestream_put_ag(ip
->i_mount
, item
->ag
);
368 TRACE_FREE(ip
->i_mount
, ip
, item
->pip
, item
->ag
,
369 xfs_filestream_peek_ag(ip
->i_mount
, item
->ag
));
372 * _xfs_filestream_update_ag() always takes a reference on the inode
373 * itself, whether it's a file or a directory. Release it here.
374 * This can result in the inode being freed and so we must
375 * not hold any inode locks when freeing filesstreams objects
376 * otherwise we can deadlock here.
381 * In the case of a regular file, _xfs_filestream_update_ag() also
382 * takes a ref on the parent inode to keep it in-core. Release that
388 /* Finally, free the memory allocated for the item. */
389 kmem_zone_free(item_zone
, item
);
393 * xfs_filestream_init() is called at xfs initialisation time to set up the
394 * memory zone that will be used for filestream data structure allocation.
397 xfs_filestream_init(void)
399 item_zone
= kmem_zone_init(sizeof(fstrm_item_t
), "fstrm_item");
407 * xfs_filestream_uninit() is called at xfs termination time to destroy the
408 * memory zone that was used for filestream data structure allocation.
411 xfs_filestream_uninit(void)
413 kmem_zone_destroy(item_zone
);
417 * xfs_filestream_mount() is called when a file system is mounted with the
418 * filestream option. It is responsible for allocating the data structures
419 * needed to track the new file system's file streams.
422 xfs_filestream_mount(
426 unsigned int lifetime
, grp_count
;
429 * The filestream timer tunable is currently fixed within the range of
430 * one second to four minutes, with five seconds being the default. The
431 * group count is somewhat arbitrary, but it'd be nice to adhere to the
432 * timer tunable to within about 10 percent. This requires at least 10
435 lifetime
= xfs_fstrm_centisecs
* 10;
438 err
= xfs_mru_cache_create(&mp
->m_filestream
, lifetime
, grp_count
,
439 xfs_fstrm_free_func
);
445 * xfs_filestream_unmount() is called when a file system that was mounted with
446 * the filestream option is unmounted. It drains the data structures created
447 * to track the file system's file streams and frees all the memory that was
451 xfs_filestream_unmount(
454 xfs_mru_cache_destroy(mp
->m_filestream
);
458 * Return the AG of the filestream the file or directory belongs to, or
459 * NULLAGNUMBER otherwise.
462 xfs_filestream_lookup_ag(
465 xfs_mru_cache_t
*cache
;
470 if (!(ip
->i_d
.di_mode
& (S_IFREG
| S_IFDIR
))) {
475 cache
= ip
->i_mount
->m_filestream
;
476 item
= xfs_mru_cache_lookup(cache
, ip
->i_ino
);
478 TRACE_LOOKUP(ip
->i_mount
, ip
, NULL
, NULLAGNUMBER
, 0);
482 ASSERT(ip
== item
->ip
);
484 ref
= xfs_filestream_peek_ag(ip
->i_mount
, ag
);
485 xfs_mru_cache_done(cache
);
487 TRACE_LOOKUP(ip
->i_mount
, ip
, item
->pip
, ag
, ref
);
492 * xfs_filestream_associate() should only be called to associate a regular file
493 * with its parent directory. Calling it with a child directory isn't
494 * appropriate because filestreams don't apply to entire directory hierarchies.
495 * Creating a file in a child directory of an existing filestream directory
496 * starts a new filestream with its own allocation group association.
498 * Returns < 0 on error, 0 if successful association occurred, > 0 if
499 * we failed to get an association because of locking issues.
502 xfs_filestream_associate(
507 xfs_mru_cache_t
*cache
;
509 xfs_agnumber_t ag
, rotorstep
, startag
;
512 ASSERT(pip
->i_d
.di_mode
& S_IFDIR
);
513 ASSERT(ip
->i_d
.di_mode
& S_IFREG
);
514 if (!(pip
->i_d
.di_mode
& S_IFDIR
) || !(ip
->i_d
.di_mode
& S_IFREG
))
518 cache
= mp
->m_filestream
;
521 * We have a problem, Houston.
523 * Taking the iolock here violates inode locking order - we already
524 * hold the ilock. Hence if we block getting this lock we may never
525 * wake. Unfortunately, that means if we can't get the lock, we're
526 * screwed in terms of getting a stream association - we can't spin
527 * waiting for the lock because someone else is waiting on the lock we
528 * hold and we cannot drop that as we are in a transaction here.
530 * Lucky for us, this inversion is not a problem because it's a
531 * directory inode that we are trying to lock here.
533 * So, if we can't get the iolock without sleeping then just give up
535 if (!xfs_ilock_nowait(pip
, XFS_IOLOCK_EXCL
))
538 /* If the parent directory is already in the cache, use its AG. */
539 item
= xfs_mru_cache_lookup(cache
, pip
->i_ino
);
541 ASSERT(item
->ip
== pip
);
543 xfs_mru_cache_done(cache
);
545 TRACE_LOOKUP(mp
, pip
, pip
, ag
, xfs_filestream_peek_ag(mp
, ag
));
546 err
= _xfs_filestream_update_ag(ip
, pip
, ag
);
552 * Set the starting AG using the rotor for inode32, otherwise
553 * use the directory inode's AG.
555 if (mp
->m_flags
& XFS_MOUNT_32BITINODES
) {
556 rotorstep
= xfs_rotorstep
;
557 startag
= (mp
->m_agfrotor
/ rotorstep
) % mp
->m_sb
.sb_agcount
;
558 mp
->m_agfrotor
= (mp
->m_agfrotor
+ 1) %
559 (mp
->m_sb
.sb_agcount
* rotorstep
);
561 startag
= XFS_INO_TO_AGNO(mp
, pip
->i_ino
);
563 /* Pick a new AG for the parent inode starting at startag. */
564 err
= _xfs_filestream_pick_ag(mp
, startag
, &ag
, 0, 0);
565 if (err
|| ag
== NULLAGNUMBER
)
568 /* Associate the parent inode with the AG. */
569 err
= _xfs_filestream_update_ag(pip
, NULL
, ag
);
573 /* Associate the file inode with the AG. */
574 err
= _xfs_filestream_update_ag(ip
, pip
, ag
);
578 TRACE_ASSOCIATE(mp
, ip
, pip
, ag
, xfs_filestream_peek_ag(mp
, ag
));
582 * If _xfs_filestream_pick_ag() returned a valid AG, remove the
583 * reference it took on it, since the file and directory will have taken
584 * their own now if they were successfully cached.
586 if (ag
!= NULLAGNUMBER
)
587 xfs_filestream_put_ag(mp
, ag
);
590 xfs_iunlock(pip
, XFS_IOLOCK_EXCL
);
595 * Pick a new allocation group for the current file and its file stream. This
596 * function is called by xfs_bmap_filestreams() with the mount point's per-ag
600 xfs_filestream_new_ag(
605 xfs_inode_t
*ip
, *pip
= NULL
;
607 xfs_mru_cache_t
*cache
;
609 fstrm_item_t
*dir
, *file
;
610 xfs_agnumber_t ag
= NULLAGNUMBER
;
614 cache
= mp
->m_filestream
;
619 * Look for the file in the cache, removing it if it's found. Doing
620 * this allows it to be held across the dir lookup that follows.
622 file
= xfs_mru_cache_remove(cache
, ip
->i_ino
);
624 ASSERT(ip
== file
->ip
);
626 /* Save the file's parent inode and old AG number for later. */
630 /* Look for the file's directory in the cache. */
631 dir
= xfs_mru_cache_lookup(cache
, pip
->i_ino
);
633 ASSERT(pip
== dir
->ip
);
636 * If the directory has already moved on to a new AG,
637 * use that AG as the new AG for the file. Don't
638 * forget to twiddle the AG refcounts to match the
641 if (dir
->ag
!= file
->ag
) {
642 xfs_filestream_put_ag(mp
, file
->ag
);
643 xfs_filestream_get_ag(mp
, dir
->ag
);
644 *agp
= file
->ag
= dir
->ag
;
647 xfs_mru_cache_done(cache
);
651 * Put the file back in the cache. If this fails, the free
652 * function needs to be called to tidy up in the same way as if
653 * the item had simply expired from the cache.
655 err
= xfs_mru_cache_insert(cache
, ip
->i_ino
, file
);
657 xfs_fstrm_free_func(ip
->i_ino
, file
);
662 * If the file's AG was moved to the directory's new AG, there's
663 * nothing more to be done.
665 if (*agp
!= NULLAGNUMBER
) {
666 TRACE_MOVEAG(mp
, ip
, pip
,
667 ag
, xfs_filestream_peek_ag(mp
, ag
),
668 *agp
, xfs_filestream_peek_ag(mp
, *agp
));
674 * If the file's parent directory is known, take its iolock in exclusive
675 * mode to prevent two sibling files from racing each other to migrate
676 * themselves and their parent to different AGs.
679 xfs_ilock(pip
, XFS_IOLOCK_EXCL
);
682 * A new AG needs to be found for the file. If the file's parent
683 * directory is also known, it will be moved to the new AG as well to
684 * ensure that files created inside it in future use the new AG.
686 ag
= (ag
== NULLAGNUMBER
) ? 0 : (ag
+ 1) % mp
->m_sb
.sb_agcount
;
687 flags
= (ap
->userdata
? XFS_PICK_USERDATA
: 0) |
688 (ap
->low
? XFS_PICK_LOWSPACE
: 0);
690 err
= _xfs_filestream_pick_ag(mp
, ag
, agp
, flags
, minlen
);
691 if (err
|| *agp
== NULLAGNUMBER
)
695 * If the file wasn't found in the file cache, then its parent directory
696 * inode isn't known. For this to have happened, the file must either
697 * be pre-existing, or it was created long enough ago that its cache
698 * entry has expired. This isn't the sort of usage that the filestreams
699 * allocator is trying to optimise, so there's no point trying to track
700 * its new AG somehow in the filestream data structures.
703 TRACE_ORPHAN(mp
, ip
, *agp
);
707 /* Associate the parent inode with the AG. */
708 err
= _xfs_filestream_update_ag(pip
, NULL
, *agp
);
712 /* Associate the file inode with the AG. */
713 err
= _xfs_filestream_update_ag(ip
, pip
, *agp
);
717 TRACE_MOVEAG(mp
, ip
, pip
, NULLAGNUMBER
, 0,
718 *agp
, xfs_filestream_peek_ag(mp
, *agp
));
722 * If _xfs_filestream_pick_ag() returned a valid AG, remove the
723 * reference it took on it, since the file and directory will have taken
724 * their own now if they were successfully cached.
726 if (*agp
!= NULLAGNUMBER
)
727 xfs_filestream_put_ag(mp
, *agp
);
732 xfs_iunlock(pip
, XFS_IOLOCK_EXCL
);
738 * Remove an association between an inode and a filestream object.
739 * Typically this is done on last close of an unlinked file.
742 xfs_filestream_deassociate(
745 xfs_mru_cache_t
*cache
= ip
->i_mount
->m_filestream
;
747 xfs_mru_cache_delete(cache
, ip
->i_ino
);