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
20 #include "xfs_bmap_btree.h"
22 #include "xfs_dinode.h"
23 #include "xfs_inode.h"
25 #include "xfs_trans.h"
27 #include "xfs_mount.h"
29 #include "xfs_bmap_util.h"
30 #include "xfs_alloc.h"
31 #include "xfs_utils.h"
32 #include "xfs_mru_cache.h"
33 #include "xfs_filestream.h"
34 #include "xfs_trace.h"
36 #ifdef XFS_FILESTREAMS_TRACE
38 ktrace_t *xfs_filestreams_trace_buf;
41 xfs_filestreams_trace(
42 xfs_mount_t *mp, /* mount point */
43 int type, /* type of trace */
44 const char *func, /* source function */
45 int line, /* source line number */
53 ktrace_enter(xfs_filestreams_trace_buf,
54 (void *)(__psint_t)(type | (line << 16)),
56 (void *)(__psunsigned_t)current_pid(),
58 (void *)(__psunsigned_t)arg0,
59 (void *)(__psunsigned_t)arg1,
60 (void *)(__psunsigned_t)arg2,
61 (void *)(__psunsigned_t)arg3,
62 (void *)(__psunsigned_t)arg4,
63 (void *)(__psunsigned_t)arg5,
64 NULL, NULL, NULL, NULL, NULL, NULL);
67 #define TRACE0(mp,t) TRACE6(mp,t,0,0,0,0,0,0)
68 #define TRACE1(mp,t,a0) TRACE6(mp,t,a0,0,0,0,0,0)
69 #define TRACE2(mp,t,a0,a1) TRACE6(mp,t,a0,a1,0,0,0,0)
70 #define TRACE3(mp,t,a0,a1,a2) TRACE6(mp,t,a0,a1,a2,0,0,0)
71 #define TRACE4(mp,t,a0,a1,a2,a3) TRACE6(mp,t,a0,a1,a2,a3,0,0)
72 #define TRACE5(mp,t,a0,a1,a2,a3,a4) TRACE6(mp,t,a0,a1,a2,a3,a4,0)
73 #define TRACE6(mp,t,a0,a1,a2,a3,a4,a5) \
74 xfs_filestreams_trace(mp, t, __func__, __LINE__, \
75 (__psunsigned_t)a0, (__psunsigned_t)a1, \
76 (__psunsigned_t)a2, (__psunsigned_t)a3, \
77 (__psunsigned_t)a4, (__psunsigned_t)a5)
79 #define TRACE_AG_SCAN(mp, ag, ag2) \
80 TRACE2(mp, XFS_FSTRM_KTRACE_AGSCAN, ag, ag2);
81 #define TRACE_AG_PICK1(mp, max_ag, maxfree) \
82 TRACE2(mp, XFS_FSTRM_KTRACE_AGPICK1, max_ag, maxfree);
83 #define TRACE_AG_PICK2(mp, ag, ag2, cnt, free, scan, flag) \
84 TRACE6(mp, XFS_FSTRM_KTRACE_AGPICK2, ag, ag2, \
85 cnt, free, scan, flag)
86 #define TRACE_UPDATE(mp, ip, ag, cnt, ag2, cnt2) \
87 TRACE5(mp, XFS_FSTRM_KTRACE_UPDATE, ip, ag, cnt, ag2, cnt2)
88 #define TRACE_FREE(mp, ip, pip, ag, cnt) \
89 TRACE4(mp, XFS_FSTRM_KTRACE_FREE, ip, pip, ag, cnt)
90 #define TRACE_LOOKUP(mp, ip, pip, ag, cnt) \
91 TRACE4(mp, XFS_FSTRM_KTRACE_ITEM_LOOKUP, ip, pip, ag, cnt)
92 #define TRACE_ASSOCIATE(mp, ip, pip, ag, cnt) \
93 TRACE4(mp, XFS_FSTRM_KTRACE_ASSOCIATE, ip, pip, ag, cnt)
94 #define TRACE_MOVEAG(mp, ip, pip, oag, ocnt, nag, ncnt) \
95 TRACE6(mp, XFS_FSTRM_KTRACE_MOVEAG, ip, pip, oag, ocnt, nag, ncnt)
96 #define TRACE_ORPHAN(mp, ip, ag) \
97 TRACE2(mp, XFS_FSTRM_KTRACE_ORPHAN, ip, ag);
101 #define TRACE_AG_SCAN(mp, ag, ag2)
102 #define TRACE_AG_PICK1(mp, max_ag, maxfree)
103 #define TRACE_AG_PICK2(mp, ag, ag2, cnt, free, scan, flag)
104 #define TRACE_UPDATE(mp, ip, ag, cnt, ag2, cnt2)
105 #define TRACE_FREE(mp, ip, pip, ag, cnt)
106 #define TRACE_LOOKUP(mp, ip, pip, ag, cnt)
107 #define TRACE_ASSOCIATE(mp, ip, pip, ag, cnt)
108 #define TRACE_MOVEAG(mp, ip, pip, oag, ocnt, nag, ncnt)
109 #define TRACE_ORPHAN(mp, ip, ag)
112 static kmem_zone_t *item_zone;
115 * Structure for associating a file or a directory with an allocation group.
116 * The parent directory pointer is only needed for files, but since there will
117 * generally be vastly more files than directories in the cache, using the same
118 * data structure simplifies the code with very little memory overhead.
120 typedef struct fstrm_item
122 xfs_agnumber_t ag; /* AG currently in use for the file/directory. */
123 xfs_inode_t *ip; /* inode self-pointer. */
124 xfs_inode_t *pip; /* Parent directory inode pointer. */
128 * Allocation group filestream associations are tracked with per-ag atomic
129 * counters. These counters allow _xfs_filestream_pick_ag() to tell whether a
130 * particular AG already has active filestreams associated with it. The mount
131 * point's m_peraglock is used to protect these counters from per-ag array
132 * re-allocation during a growfs operation. When xfs_growfs_data_private() is
133 * about to reallocate the array, it calls xfs_filestream_flush() with the
134 * m_peraglock held in write mode.
136 * Since xfs_mru_cache_flush() guarantees that all the free functions for all
137 * the cache elements have finished executing before it returns, it's safe for
138 * the free functions to use the atomic counters without m_peraglock protection.
139 * This allows the implementation of xfs_fstrm_free_func() to be agnostic about
140 * whether it was called with the m_peraglock held in read mode, write mode or
141 * not held at all. The race condition this addresses is the following:
143 * - The work queue scheduler fires and pulls a filestream directory cache
144 * element off the LRU end of the cache for deletion, then gets pre-empted.
145 * - A growfs operation grabs the m_peraglock in write mode, flushes all the
146 * remaining items from the cache and reallocates the mount point's per-ag
147 * array, resetting all the counters to zero.
148 * - The work queue thread resumes and calls the free function for the element
149 * it started cleaning up earlier. In the process it decrements the
150 * filestreams counter for an AG that now has no references.
152 * With a shrinkfs feature, the above scenario could panic the system.
154 * All other uses of the following macros should be protected by either the
155 * m_peraglock held in read mode, or the cache's internal locking exposed by the
156 * interval between a call to xfs_mru_cache_lookup() and a call to
157 * xfs_mru_cache_done(). In addition, the m_peraglock must be held in read mode
158 * when new elements are added to the cache.
160 * Combined, these locking rules ensure that no associations will ever exist in
161 * the cache that reference per-ag array elements that have since been
165 xfs_filestream_peek_ag(
169 struct xfs_perag *pag;
172 pag = xfs_perag_get(mp, agno);
173 ret = atomic_read(&pag->pagf_fstrms);
179 xfs_filestream_get_ag(
183 struct xfs_perag *pag;
186 pag = xfs_perag_get(mp, agno);
187 ret = atomic_inc_return(&pag->pagf_fstrms);
193 xfs_filestream_put_ag(
197 struct xfs_perag *pag;
199 pag = xfs_perag_get(mp, agno);
200 atomic_dec(&pag->pagf_fstrms);
205 * Scan the AGs starting at startag looking for an AG that isn't in use and has
206 * at least minlen blocks free.
209 _xfs_filestream_pick_ag(
211 xfs_agnumber_t startag,
216 int streams, max_streams;
217 int err, trylock, nscan;
218 xfs_extlen_t longest, free, minfree, maxfree = 0;
219 xfs_agnumber_t ag, max_ag = NULLAGNUMBER;
220 struct xfs_perag *pag;
222 /* 2% of an AG's blocks must be free for it to be chosen. */
223 minfree = mp->m_sb.sb_agblocks / 50;
228 /* For the first pass, don't sleep trying to init the per-AG. */
229 trylock = XFS_ALLOC_FLAG_TRYLOCK;
231 for (nscan = 0; 1; nscan++) {
232 pag = xfs_perag_get(mp, ag);
233 TRACE_AG_SCAN(mp, ag, atomic_read(&pag->pagf_fstrms));
235 if (!pag->pagf_init) {
236 err = xfs_alloc_pagf_init(mp, NULL, ag, trylock);
237 if (err && !trylock) {
243 /* Might fail sometimes during the 1st pass with trylock set. */
247 /* Keep track of the AG with the most free blocks. */
248 if (pag->pagf_freeblks > maxfree) {
249 maxfree = pag->pagf_freeblks;
250 max_streams = atomic_read(&pag->pagf_fstrms);
255 * The AG reference count does two things: it enforces mutual
256 * exclusion when examining the suitability of an AG in this
257 * loop, and it guards against two filestreams being established
258 * in the same AG as each other.
260 if (xfs_filestream_get_ag(mp, ag) > 1) {
261 xfs_filestream_put_ag(mp, ag);
265 longest = xfs_alloc_longest_free_extent(mp, pag);
266 if (((minlen && longest >= minlen) ||
267 (!minlen && pag->pagf_freeblks >= minfree)) &&
268 (!pag->pagf_metadata || !(flags & XFS_PICK_USERDATA) ||
269 (flags & XFS_PICK_LOWSPACE))) {
271 /* Break out, retaining the reference on the AG. */
272 free = pag->pagf_freeblks;
273 streams = atomic_read(&pag->pagf_fstrms);
279 /* Drop the reference on this AG, it's not usable. */
280 xfs_filestream_put_ag(mp, ag);
283 /* Move to the next AG, wrapping to AG 0 if necessary. */
284 if (++ag >= mp->m_sb.sb_agcount)
287 /* If a full pass of the AGs hasn't been done yet, continue. */
291 /* Allow sleeping in xfs_alloc_pagf_init() on the 2nd pass. */
297 /* Finally, if lowspace wasn't set, set it for the 3rd pass. */
298 if (!(flags & XFS_PICK_LOWSPACE)) {
299 flags |= XFS_PICK_LOWSPACE;
304 * Take the AG with the most free space, regardless of whether
305 * it's already in use by another filestream.
307 if (max_ag != NULLAGNUMBER) {
308 xfs_filestream_get_ag(mp, max_ag);
309 TRACE_AG_PICK1(mp, max_ag, maxfree);
310 streams = max_streams;
316 /* take AG 0 if none matched */
317 TRACE_AG_PICK1(mp, max_ag, maxfree);
322 TRACE_AG_PICK2(mp, startag, *agp, streams, free, nscan, flags);
328 * Set the allocation group number for a file or a directory, updating inode
329 * references and per-AG references as appropriate.
332 _xfs_filestream_update_ag(
339 xfs_mru_cache_t *cache;
341 xfs_agnumber_t old_ag;
342 xfs_inode_t *old_pip;
345 * Either ip is a regular file and pip is a directory, or ip is a
346 * directory and pip is NULL.
348 ASSERT(ip && ((S_ISREG(ip->i_d.di_mode) && pip &&
349 S_ISDIR(pip->i_d.di_mode)) ||
350 (S_ISDIR(ip->i_d.di_mode) && !pip)));
353 cache = mp->m_filestream;
355 item = xfs_mru_cache_lookup(cache, ip->i_ino);
357 ASSERT(item->ip == ip);
362 xfs_mru_cache_done(cache);
365 * If the AG has changed, drop the old ref and take a new one,
366 * effectively transferring the reference from old to new AG.
369 xfs_filestream_put_ag(mp, old_ag);
370 xfs_filestream_get_ag(mp, ag);
374 * If ip is a file and its pip has changed, drop the old ref and
377 if (pip && pip != old_pip) {
382 TRACE_UPDATE(mp, ip, old_ag, xfs_filestream_peek_ag(mp, old_ag),
383 ag, xfs_filestream_peek_ag(mp, ag));
387 item = kmem_zone_zalloc(item_zone, KM_MAYFAIL);
395 err = xfs_mru_cache_insert(cache, ip->i_ino, item);
397 kmem_zone_free(item_zone, item);
401 /* Take a reference on the AG. */
402 xfs_filestream_get_ag(mp, ag);
405 * Take a reference on the inode itself regardless of whether it's a
406 * regular file or a directory.
411 * In the case of a regular file, take a reference on the parent inode
412 * as well to ensure it remains in-core.
417 TRACE_UPDATE(mp, ip, ag, xfs_filestream_peek_ag(mp, ag),
418 ag, xfs_filestream_peek_ag(mp, ag));
423 /* xfs_fstrm_free_func(): callback for freeing cached stream items. */
429 fstrm_item_t *item = (fstrm_item_t *)data;
430 xfs_inode_t *ip = item->ip;
432 ASSERT(ip->i_ino == ino);
434 xfs_iflags_clear(ip, XFS_IFILESTREAM);
436 /* Drop the reference taken on the AG when the item was added. */
437 xfs_filestream_put_ag(ip->i_mount, item->ag);
439 TRACE_FREE(ip->i_mount, ip, item->pip, item->ag,
440 xfs_filestream_peek_ag(ip->i_mount, item->ag));
443 * _xfs_filestream_update_ag() always takes a reference on the inode
444 * itself, whether it's a file or a directory. Release it here.
445 * This can result in the inode being freed and so we must
446 * not hold any inode locks when freeing filesstreams objects
447 * otherwise we can deadlock here.
452 * In the case of a regular file, _xfs_filestream_update_ag() also
453 * takes a ref on the parent inode to keep it in-core. Release that
459 /* Finally, free the memory allocated for the item. */
460 kmem_zone_free(item_zone, item);
464 * xfs_filestream_init() is called at xfs initialisation time to set up the
465 * memory zone that will be used for filestream data structure allocation.
468 xfs_filestream_init(void)
470 item_zone = kmem_zone_init(sizeof(fstrm_item_t), "fstrm_item");
478 * xfs_filestream_uninit() is called at xfs termination time to destroy the
479 * memory zone that was used for filestream data structure allocation.
482 xfs_filestream_uninit(void)
484 kmem_zone_destroy(item_zone);
488 * xfs_filestream_mount() is called when a file system is mounted with the
489 * filestream option. It is responsible for allocating the data structures
490 * needed to track the new file system's file streams.
493 xfs_filestream_mount(
497 unsigned int lifetime, grp_count;
500 * The filestream timer tunable is currently fixed within the range of
501 * one second to four minutes, with five seconds being the default. The
502 * group count is somewhat arbitrary, but it'd be nice to adhere to the
503 * timer tunable to within about 10 percent. This requires at least 10
506 lifetime = xfs_fstrm_centisecs * 10;
509 err = xfs_mru_cache_create(&mp->m_filestream, lifetime, grp_count,
510 xfs_fstrm_free_func);
516 * xfs_filestream_unmount() is called when a file system that was mounted with
517 * the filestream option is unmounted. It drains the data structures created
518 * to track the file system's file streams and frees all the memory that was
522 xfs_filestream_unmount(
525 xfs_mru_cache_destroy(mp->m_filestream);
529 * Return the AG of the filestream the file or directory belongs to, or
530 * NULLAGNUMBER otherwise.
533 xfs_filestream_lookup_ag(
536 xfs_mru_cache_t *cache;
541 if (!S_ISREG(ip->i_d.di_mode) && !S_ISDIR(ip->i_d.di_mode)) {
546 cache = ip->i_mount->m_filestream;
547 item = xfs_mru_cache_lookup(cache, ip->i_ino);
549 TRACE_LOOKUP(ip->i_mount, ip, NULL, NULLAGNUMBER, 0);
553 ASSERT(ip == item->ip);
555 ref = xfs_filestream_peek_ag(ip->i_mount, ag);
556 xfs_mru_cache_done(cache);
558 TRACE_LOOKUP(ip->i_mount, ip, item->pip, ag, ref);
563 * xfs_filestream_associate() should only be called to associate a regular file
564 * with its parent directory. Calling it with a child directory isn't
565 * appropriate because filestreams don't apply to entire directory hierarchies.
566 * Creating a file in a child directory of an existing filestream directory
567 * starts a new filestream with its own allocation group association.
569 * Returns < 0 on error, 0 if successful association occurred, > 0 if
570 * we failed to get an association because of locking issues.
573 xfs_filestream_associate(
578 xfs_mru_cache_t *cache;
580 xfs_agnumber_t ag, rotorstep, startag;
583 ASSERT(S_ISDIR(pip->i_d.di_mode));
584 ASSERT(S_ISREG(ip->i_d.di_mode));
585 if (!S_ISDIR(pip->i_d.di_mode) || !S_ISREG(ip->i_d.di_mode))
589 cache = mp->m_filestream;
592 * We have a problem, Houston.
594 * Taking the iolock here violates inode locking order - we already
595 * hold the ilock. Hence if we block getting this lock we may never
596 * wake. Unfortunately, that means if we can't get the lock, we're
597 * screwed in terms of getting a stream association - we can't spin
598 * waiting for the lock because someone else is waiting on the lock we
599 * hold and we cannot drop that as we are in a transaction here.
601 * Lucky for us, this inversion is not a problem because it's a
602 * directory inode that we are trying to lock here.
604 * So, if we can't get the iolock without sleeping then just give up
606 if (!xfs_ilock_nowait(pip, XFS_IOLOCK_EXCL))
609 /* If the parent directory is already in the cache, use its AG. */
610 item = xfs_mru_cache_lookup(cache, pip->i_ino);
612 ASSERT(item->ip == pip);
614 xfs_mru_cache_done(cache);
616 TRACE_LOOKUP(mp, pip, pip, ag, xfs_filestream_peek_ag(mp, ag));
617 err = _xfs_filestream_update_ag(ip, pip, ag);
623 * Set the starting AG using the rotor for inode32, otherwise
624 * use the directory inode's AG.
626 if (mp->m_flags & XFS_MOUNT_32BITINODES) {
627 rotorstep = xfs_rotorstep;
628 startag = (mp->m_agfrotor / rotorstep) % mp->m_sb.sb_agcount;
629 mp->m_agfrotor = (mp->m_agfrotor + 1) %
630 (mp->m_sb.sb_agcount * rotorstep);
632 startag = XFS_INO_TO_AGNO(mp, pip->i_ino);
634 /* Pick a new AG for the parent inode starting at startag. */
635 err = _xfs_filestream_pick_ag(mp, startag, &ag, 0, 0);
636 if (err || ag == NULLAGNUMBER)
639 /* Associate the parent inode with the AG. */
640 err = _xfs_filestream_update_ag(pip, NULL, ag);
644 /* Associate the file inode with the AG. */
645 err = _xfs_filestream_update_ag(ip, pip, ag);
649 TRACE_ASSOCIATE(mp, ip, pip, ag, xfs_filestream_peek_ag(mp, ag));
653 * If _xfs_filestream_pick_ag() returned a valid AG, remove the
654 * reference it took on it, since the file and directory will have taken
655 * their own now if they were successfully cached.
657 if (ag != NULLAGNUMBER)
658 xfs_filestream_put_ag(mp, ag);
661 xfs_iunlock(pip, XFS_IOLOCK_EXCL);
666 * Pick a new allocation group for the current file and its file stream. This
667 * function is called by xfs_bmap_filestreams() with the mount point's per-ag
671 xfs_filestream_new_ag(
672 struct xfs_bmalloca *ap,
676 xfs_inode_t *ip, *pip = NULL;
678 xfs_mru_cache_t *cache;
680 fstrm_item_t *dir, *file;
681 xfs_agnumber_t ag = NULLAGNUMBER;
685 cache = mp->m_filestream;
690 * Look for the file in the cache, removing it if it's found. Doing
691 * this allows it to be held across the dir lookup that follows.
693 file = xfs_mru_cache_remove(cache, ip->i_ino);
695 ASSERT(ip == file->ip);
697 /* Save the file's parent inode and old AG number for later. */
701 /* Look for the file's directory in the cache. */
702 dir = xfs_mru_cache_lookup(cache, pip->i_ino);
704 ASSERT(pip == dir->ip);
707 * If the directory has already moved on to a new AG,
708 * use that AG as the new AG for the file. Don't
709 * forget to twiddle the AG refcounts to match the
712 if (dir->ag != file->ag) {
713 xfs_filestream_put_ag(mp, file->ag);
714 xfs_filestream_get_ag(mp, dir->ag);
715 *agp = file->ag = dir->ag;
718 xfs_mru_cache_done(cache);
722 * Put the file back in the cache. If this fails, the free
723 * function needs to be called to tidy up in the same way as if
724 * the item had simply expired from the cache.
726 err = xfs_mru_cache_insert(cache, ip->i_ino, file);
728 xfs_fstrm_free_func(ip->i_ino, file);
733 * If the file's AG was moved to the directory's new AG, there's
734 * nothing more to be done.
736 if (*agp != NULLAGNUMBER) {
737 TRACE_MOVEAG(mp, ip, pip,
738 ag, xfs_filestream_peek_ag(mp, ag),
739 *agp, xfs_filestream_peek_ag(mp, *agp));
745 * If the file's parent directory is known, take its iolock in exclusive
746 * mode to prevent two sibling files from racing each other to migrate
747 * themselves and their parent to different AGs.
749 * Note that we lock the parent directory iolock inside the child
750 * iolock here. That's fine as we never hold both parent and child
751 * iolock in any other place. This is different from the ilock,
752 * which requires locking of the child after the parent for namespace
756 xfs_ilock(pip, XFS_IOLOCK_EXCL | XFS_IOLOCK_PARENT);
759 * A new AG needs to be found for the file. If the file's parent
760 * directory is also known, it will be moved to the new AG as well to
761 * ensure that files created inside it in future use the new AG.
763 ag = (ag == NULLAGNUMBER) ? 0 : (ag + 1) % mp->m_sb.sb_agcount;
764 flags = (ap->userdata ? XFS_PICK_USERDATA : 0) |
765 (ap->flist->xbf_low ? XFS_PICK_LOWSPACE : 0);
767 err = _xfs_filestream_pick_ag(mp, ag, agp, flags, minlen);
768 if (err || *agp == NULLAGNUMBER)
772 * If the file wasn't found in the file cache, then its parent directory
773 * inode isn't known. For this to have happened, the file must either
774 * be pre-existing, or it was created long enough ago that its cache
775 * entry has expired. This isn't the sort of usage that the filestreams
776 * allocator is trying to optimise, so there's no point trying to track
777 * its new AG somehow in the filestream data structures.
780 TRACE_ORPHAN(mp, ip, *agp);
784 /* Associate the parent inode with the AG. */
785 err = _xfs_filestream_update_ag(pip, NULL, *agp);
789 /* Associate the file inode with the AG. */
790 err = _xfs_filestream_update_ag(ip, pip, *agp);
794 TRACE_MOVEAG(mp, ip, pip, NULLAGNUMBER, 0,
795 *agp, xfs_filestream_peek_ag(mp, *agp));
799 * If _xfs_filestream_pick_ag() returned a valid AG, remove the
800 * reference it took on it, since the file and directory will have taken
801 * their own now if they were successfully cached.
803 if (*agp != NULLAGNUMBER)
804 xfs_filestream_put_ag(mp, *agp);
809 xfs_iunlock(pip, XFS_IOLOCK_EXCL);
815 * Remove an association between an inode and a filestream object.
816 * Typically this is done on last close of an unlinked file.
819 xfs_filestream_deassociate(
822 xfs_mru_cache_t *cache = ip->i_mount->m_filestream;
824 xfs_mru_cache_delete(cache, ip->i_ino);