xfs: rewrite the filestream allocator using the dentry cache

In Linux we will always be able to find a parent inode for file that are
undergoing I/O.  Use this to simply the file stream allocator by only
keeping track of parent inodes.

Signed-off-by: Christoph Hellwig <hch@lst.de>

1 files changed, 165 insertions, 495 deletions

diff --git a/fs/xfs/xfs_filestream.c b/fs/xfs/xfs_filestream.c
index c422110c7c43..ff6f90215c8a 100644
--- a/fs/xfs/xfs_filestream.c
+++ b/fs/xfs/xfs_filestream.c
@@ -1,5 +1,6 @@
 /*
  * Copyright (c) 2006-2007 Silicon Graphics, Inc.
+ * Copyright (c) 2014 Christoph Hellwig.
  * All Rights Reserved.
  *
  * This program is free software; you can redistribute it and/or
@@ -32,101 +33,28 @@
 #include "xfs_filestream.h"
 #include "xfs_trace.h"
 
-#ifdef XFS_FILESTREAMS_TRACE
-
-ktrace_t *xfs_filestreams_trace_buf;
-
-STATIC void
-xfs_filestreams_trace(
-	xfs_mount_t	*mp,	/* mount point */
-	int		type,	/* type of trace */
-	const char	*func,	/* source function */
-	int		line,	/* source line number */
-	__psunsigned_t	arg0,
-	__psunsigned_t	arg1,
-	__psunsigned_t	arg2,
-	__psunsigned_t	arg3,
-	__psunsigned_t	arg4,
-	__psunsigned_t	arg5)
-{
-	ktrace_enter(xfs_filestreams_trace_buf,
-		(void *)(__psint_t)(type | (line << 16)),
-		(void *)func,
-		(void *)(__psunsigned_t)current_pid(),
-		(void *)mp,
-		(void *)(__psunsigned_t)arg0,
-		(void *)(__psunsigned_t)arg1,
-		(void *)(__psunsigned_t)arg2,
-		(void *)(__psunsigned_t)arg3,
-		(void *)(__psunsigned_t)arg4,
-		(void *)(__psunsigned_t)arg5,
-		NULL, NULL, NULL, NULL, NULL, NULL);
-}
-
-#define TRACE0(mp,t)			TRACE6(mp,t,0,0,0,0,0,0)
-#define TRACE1(mp,t,a0)			TRACE6(mp,t,a0,0,0,0,0,0)
-#define TRACE2(mp,t,a0,a1)		TRACE6(mp,t,a0,a1,0,0,0,0)
-#define TRACE3(mp,t,a0,a1,a2)		TRACE6(mp,t,a0,a1,a2,0,0,0)
-#define TRACE4(mp,t,a0,a1,a2,a3)	TRACE6(mp,t,a0,a1,a2,a3,0,0)
-#define TRACE5(mp,t,a0,a1,a2,a3,a4)	TRACE6(mp,t,a0,a1,a2,a3,a4,0)
-#define TRACE6(mp,t,a0,a1,a2,a3,a4,a5) \
-	xfs_filestreams_trace(mp, t, __func__, __LINE__, \
-				(__psunsigned_t)a0, (__psunsigned_t)a1, \
-				(__psunsigned_t)a2, (__psunsigned_t)a3, \
-				(__psunsigned_t)a4, (__psunsigned_t)a5)
-
-#define TRACE_AG_SCAN(mp, ag, ag2) \
-		TRACE2(mp, XFS_FSTRM_KTRACE_AGSCAN, ag, ag2);
-#define TRACE_AG_PICK1(mp, max_ag, maxfree) \
-		TRACE2(mp, XFS_FSTRM_KTRACE_AGPICK1, max_ag, maxfree);
-#define TRACE_AG_PICK2(mp, ag, ag2, cnt, free, scan, flag) \
-		TRACE6(mp, XFS_FSTRM_KTRACE_AGPICK2, ag, ag2, \
-			 cnt, free, scan, flag)
-#define TRACE_UPDATE(mp, ip, ag, cnt, ag2, cnt2) \
-		TRACE5(mp, XFS_FSTRM_KTRACE_UPDATE, ip, ag, cnt, ag2, cnt2)
-#define TRACE_FREE(mp, ip, pip, ag, cnt) \
-		TRACE4(mp, XFS_FSTRM_KTRACE_FREE, ip, pip, ag, cnt)
-#define TRACE_LOOKUP(mp, ip, pip, ag, cnt) \
-		TRACE4(mp, XFS_FSTRM_KTRACE_ITEM_LOOKUP, ip, pip, ag, cnt)
-#define TRACE_ASSOCIATE(mp, ip, pip, ag, cnt) \
-		TRACE4(mp, XFS_FSTRM_KTRACE_ASSOCIATE, ip, pip, ag, cnt)
-#define TRACE_MOVEAG(mp, ip, pip, oag, ocnt, nag, ncnt) \
-		TRACE6(mp, XFS_FSTRM_KTRACE_MOVEAG, ip, pip, oag, ocnt, nag, ncnt)
-#define TRACE_ORPHAN(mp, ip, ag) \
-		TRACE2(mp, XFS_FSTRM_KTRACE_ORPHAN, ip, ag);
-
-
-#else
 #define TRACE_AG_SCAN(mp, ag, ag2)
 #define TRACE_AG_PICK1(mp, max_ag, maxfree)
 #define TRACE_AG_PICK2(mp, ag, ag2, cnt, free, scan, flag)
-#define TRACE_UPDATE(mp, ip, ag, cnt, ag2, cnt2)
 #define TRACE_FREE(mp, ip, pip, ag, cnt)
 #define TRACE_LOOKUP(mp, ip, pip, ag, cnt)
-#define TRACE_ASSOCIATE(mp, ip, pip, ag, cnt)
-#define TRACE_MOVEAG(mp, ip, pip, oag, ocnt, nag, ncnt)
-#define TRACE_ORPHAN(mp, ip, ag)
-#endif
 
 static kmem_zone_t *item_zone;
 
-/*
- * Structure for associating a file or a directory with an allocation group.
- * The parent directory pointer is only needed for files, but since there will
- * generally be vastly more files than directories in the cache, using the same
- * data structure simplifies the code with very little memory overhead.
- */
-typedef struct fstrm_item
-{
-	struct xfs_mru_cache_elem mru;
-	xfs_agnumber_t	ag;	/* AG currently in use for the file/directory. */
-	xfs_inode_t	*ip;	/* inode self-pointer. */
-	xfs_inode_t	*pip;	/* Parent directory inode pointer. */
-} fstrm_item_t;
+struct xfs_fstrm_item {
+	struct xfs_mru_cache_elem	mru;
+	struct xfs_inode		*ip;
+	xfs_agnumber_t			ag; /* AG in use for this directory */
+};
+
+enum xfs_fstrm_alloc {
+	XFS_PICK_USERDATA = 1,
+	XFS_PICK_LOWSPACE = 2,
+};
 
 /*
  * Allocation group filestream associations are tracked with per-ag atomic
- * counters.  These counters allow _xfs_filestream_pick_ag() to tell whether a
+ * counters.  These counters allow xfs_filestream_pick_ag() to tell whether a
  * particular AG already has active filestreams associated with it. The mount
  * point's m_peraglock is used to protect these counters from per-ag array
  * re-allocation during a growfs operation.  When xfs_growfs_data_private() is
@@ -201,23 +129,42 @@ xfs_filestream_put_ag(
 	xfs_perag_put(pag);
 }
 
+static void
+xfs_fstrm_free_func(
+	struct xfs_mru_cache_elem *mru)
+{
+	struct xfs_fstrm_item	*item =
+		container_of(mru, struct xfs_fstrm_item, mru);
+
+	xfs_filestream_put_ag(item->ip->i_mount, item->ag);
+
+	TRACE_FREE(mp, ip, NULL, item->ag,
+		   xfs_filestream_peek_ag(mp, item->ag));
+
+	kmem_zone_free(item_zone, item);
+}
+
 /*
  * Scan the AGs starting at startag looking for an AG that isn't in use and has
  * at least minlen blocks free.
  */
 static int
-_xfs_filestream_pick_ag(
-	xfs_mount_t	*mp,
-	xfs_agnumber_t	startag,
-	xfs_agnumber_t	*agp,
-	int		flags,
-	xfs_extlen_t	minlen)
+xfs_filestream_pick_ag(
+	struct xfs_inode	*ip,
+	xfs_agnumber_t		startag,
+	xfs_agnumber_t		*agp,
+	int			flags,
+	xfs_extlen_t		minlen)
 {
-	int		streams, max_streams;
-	int		err, trylock, nscan;
-	xfs_extlen_t	longest, free, minfree, maxfree = 0;
-	xfs_agnumber_t	ag, max_ag = NULLAGNUMBER;
-	struct xfs_perag *pag;
+	struct xfs_mount	*mp = ip->i_mount;
+	struct xfs_fstrm_item	*item;
+	struct xfs_perag	*pag;
+	xfs_extlen_t		longest, free, minfree, maxfree = 0;
+	xfs_agnumber_t		ag, max_ag = NULLAGNUMBER;
+	int			streams, max_streams;
+	int			err, trylock, nscan;
+
+	ASSERT(S_ISDIR(ip->i_d.di_mode));
 
 	/* 2% of an AG's blocks must be free for it to be chosen. */
 	minfree = mp->m_sb.sb_agblocks / 50;
@@ -321,205 +268,55 @@ next_ag:
 
 	TRACE_AG_PICK2(mp, startag, *agp, streams, free, nscan, flags);
 
-	return 0;
-}
-
-/*
- * Set the allocation group number for a file or a directory, updating inode
- * references and per-AG references as appropriate.
- */
-static int
-_xfs_filestream_update_ag(
-	xfs_inode_t	*ip,
-	xfs_inode_t	*pip,
-	xfs_agnumber_t	ag)
-{
-	int		err = 0;
-	xfs_mount_t	*mp;
-	fstrm_item_t	*item;
-	xfs_agnumber_t	old_ag;
-	xfs_inode_t	*old_pip;
-	struct xfs_mru_cache_elem *mru;
-
-	/*
-	 * Either ip is a regular file and pip is a directory, or ip is a
-	 * directory and pip is NULL.
-	 */
-	ASSERT(ip && ((S_ISREG(ip->i_d.di_mode) && pip &&
-	               S_ISDIR(pip->i_d.di_mode)) ||
-	              (S_ISDIR(ip->i_d.di_mode) && !pip)));
-
-	mp = ip->i_mount;
-
-	mru = xfs_mru_cache_lookup(mp->m_filestream, ip->i_ino);
-	if (mru) {
-		item = container_of(mru, fstrm_item_t, mru);
-
-		ASSERT(item->ip == ip);
-		old_ag = item->ag;
-		item->ag = ag;
-		old_pip = item->pip;
-		item->pip = pip;
-		xfs_mru_cache_done(mp->m_filestream);
-
-		/*
-		 * If the AG has changed, drop the old ref and take a new one,
-		 * effectively transferring the reference from old to new AG.
-		 */
-		if (ag != old_ag) {
-			xfs_filestream_put_ag(mp, old_ag);
-			xfs_filestream_get_ag(mp, ag);
-		}
-
-		/*
-		 * If ip is a file and its pip has changed, drop the old ref and
-		 * take a new one.
-		 */
-		if (pip && pip != old_pip) {
-			IRELE(old_pip);
-			IHOLD(pip);
-		}
-
-		TRACE_UPDATE(mp, ip, old_ag, xfs_filestream_peek_ag(mp, old_ag),
-				ag, xfs_filestream_peek_ag(mp, ag));
+	if (*agp == NULLAGNUMBER)
 		return 0;
-	}
 
+	err = ENOMEM;
 	item = kmem_zone_zalloc(item_zone, KM_MAYFAIL);
 	if (!item)
-		return ENOMEM;
+		goto out_put_ag;
 
-	item->ag = ag;
+	item->ag = *agp;
 	item->ip = ip;
-	item->pip = pip;
 
 	err = xfs_mru_cache_insert(mp->m_filestream, ip->i_ino, &item->mru);
 	if (err) {
-		kmem_zone_free(item_zone, item);
-		return err;
+		if (err == EEXIST)
+			err = 0;
+		goto out_free_item;
 	}
 
-	/* Take a reference on the AG. */
-	xfs_filestream_get_ag(mp, ag);
-
-	/*
-	 * Take a reference on the inode itself regardless of whether it's a
-	 * regular file or a directory.
-	 */
-	IHOLD(ip);
-
-	/*
-	 * In the case of a regular file, take a reference on the parent inode
-	 * as well to ensure it remains in-core.
-	 */
-	if (pip)
-		IHOLD(pip);
-
-	TRACE_UPDATE(mp, ip, ag, xfs_filestream_peek_ag(mp, ag),
-			ag, xfs_filestream_peek_ag(mp, ag));
-
 	return 0;
-}
-
-/* xfs_fstrm_free_func(): callback for freeing cached stream items. */
-STATIC void
-xfs_fstrm_free_func(
-	struct xfs_mru_cache_elem *mru)
-{
-	fstrm_item_t	*item =
-		container_of(mru, fstrm_item_t, mru);
-	xfs_inode_t	*ip = item->ip;
-
-	/* Drop the reference taken on the AG when the item was added. */
-	xfs_filestream_put_ag(ip->i_mount, item->ag);
-
-	TRACE_FREE(ip->i_mount, ip, item->pip, item->ag,
-		xfs_filestream_peek_ag(ip->i_mount, item->ag));
-
-	/*
-	 * _xfs_filestream_update_ag() always takes a reference on the inode
-	 * itself, whether it's a file or a directory.  Release it here.
-	 * This can result in the inode being freed and so we must
-	 * not hold any inode locks when freeing filesstreams objects
-	 * otherwise we can deadlock here.
-	 */
-	IRELE(ip);
 
-	/*
-	 * In the case of a regular file, _xfs_filestream_update_ag() also
-	 * takes a ref on the parent inode to keep it in-core.  Release that
-	 * too.
-	 */
-	if (item->pip)
-		IRELE(item->pip);
-
-	/* Finally, free the memory allocated for the item. */
+out_free_item:
 	kmem_zone_free(item_zone, item);
+out_put_ag:
+	xfs_filestream_put_ag(mp, *agp);
+	return err;
 }
 
-/*
- * xfs_filestream_init() is called at xfs initialisation time to set up the
- * memory zone that will be used for filestream data structure allocation.
- */
-int
-xfs_filestream_init(void)
-{
-	item_zone = kmem_zone_init(sizeof(fstrm_item_t), "fstrm_item");
-	if (!item_zone)
-		return -ENOMEM;
-
-	return 0;
-}
-
-/*
- * xfs_filestream_uninit() is called at xfs termination time to destroy the
- * memory zone that was used for filestream data structure allocation.
- */
-void
-xfs_filestream_uninit(void)
+static struct xfs_inode *
+xfs_filestream_get_parent(
+	struct xfs_inode	*ip)
 {
-	kmem_zone_destroy(item_zone);
-}
+	struct inode		*inode = VFS_I(ip), *dir = NULL;
+	struct dentry		*dentry, *parent;
 
-/*
- * xfs_filestream_mount() is called when a file system is mounted with the
- * filestream option.  It is responsible for allocating the data structures
- * needed to track the new file system's file streams.
- */
-int
-xfs_filestream_mount(
-	xfs_mount_t	*mp)
-{
-	int		err;
-	unsigned int	lifetime, grp_count;
+	dentry = d_find_alias(inode);
+	if (!dentry)
+		goto out;
 
-	/*
-	 * The filestream timer tunable is currently fixed within the range of
-	 * one second to four minutes, with five seconds being the default.  The
-	 * group count is somewhat arbitrary, but it'd be nice to adhere to the
-	 * timer tunable to within about 10 percent.  This requires at least 10
-	 * groups.
-	 */
-	lifetime  = xfs_fstrm_centisecs * 10;
-	grp_count = 10;
+	parent = dget_parent(dentry);
+	if (!parent)
+		goto out_dput;
 
-	err = xfs_mru_cache_create(&mp->m_filestream, lifetime, grp_count,
-	                     xfs_fstrm_free_func);
+	dir = igrab(parent->d_inode);
+	dput(parent);
 
-	return err;
-}
-
-/*
- * xfs_filestream_unmount() is called when a file system that was mounted with
- * the filestream option is unmounted.  It drains the data structures created
- * to track the file system's file streams and frees all the memory that was
- * allocated.
- */
-void
-xfs_filestream_unmount(
-	xfs_mount_t	*mp)
-{
-	xfs_mru_cache_destroy(mp->m_filestream);
+out_dput:
+	dput(dentry);
+out:
+	return dir ? XFS_I(dir) : NULL;
 }
 
 /*
@@ -528,94 +325,61 @@ xfs_filestream_unmount(
  */
 xfs_agnumber_t
 xfs_filestream_lookup_ag(
-	xfs_inode_t	*ip)
+	struct xfs_inode	*ip)
 {
-	struct xfs_mount *mp = ip->i_mount;
+	struct xfs_mount	*mp = ip->i_mount;
+	struct xfs_fstrm_item	*item;
+	struct xfs_inode	*pip = NULL;
+	xfs_agnumber_t		ag = NULLAGNUMBER;
+	int			ref = 0;
 	struct xfs_mru_cache_elem *mru;
-	fstrm_item_t	*item;
-	xfs_agnumber_t	ag;
-	int		ref;
 
-	if (!S_ISREG(ip->i_d.di_mode) && !S_ISDIR(ip->i_d.di_mode)) {
-		ASSERT(0);
-		return NULLAGNUMBER;
-	}
+	ASSERT(S_ISREG(ip->i_d.di_mode));
 
-	mru = xfs_mru_cache_lookup(mp->m_filestream, ip->i_ino);
-	if (!mru) {
-		TRACE_LOOKUP(ip->i_mount, ip, NULL, NULLAGNUMBER, 0);
-		return NULLAGNUMBER;
-	}
+	pip = xfs_filestream_get_parent(ip);
+	if (!pip)
+		goto out;
+
+	mru = xfs_mru_cache_lookup(mp->m_filestream, pip->i_ino);
+	if (!mru)
+		goto out;
+
+	item = container_of(mru, struct xfs_fstrm_item, mru);
 
-	item = container_of(mru, fstrm_item_t, mru);
-	ASSERT(ip == item->ip);
 	ag = item->ag;
-	ref = xfs_filestream_peek_ag(ip->i_mount, ag);
 	xfs_mru_cache_done(mp->m_filestream);
 
-	TRACE_LOOKUP(ip->i_mount, ip, item->pip, ag, ref);
+	ref = xfs_filestream_peek_ag(ip->i_mount, ag);
+out:
+	TRACE_LOOKUP(mp, ip, pip, ag, ref);
+	IRELE(pip);
 	return ag;
 }
 
 /*
- * xfs_filestream_associate() should only be called to associate a regular file
- * with its parent directory.  Calling it with a child directory isn't
- * appropriate because filestreams don't apply to entire directory hierarchies.
- * Creating a file in a child directory of an existing filestream directory
- * starts a new filestream with its own allocation group association.
+ * Make sure a directory has a filestream associated with it.
  *
- * Returns < 0 on error, 0 if successful association occurred, > 0 if
- * we failed to get an association because of locking issues.
+ * This is called when creating regular files in an directory that has
+ * filestreams enabled, so that a stream is ready by the time we need it
+ * in the allocator for the files inside the directory.
  */
 int
 xfs_filestream_associate(
-	xfs_inode_t	*pip,
-	xfs_inode_t	*ip)
+	struct xfs_inode	*pip)
 {
+	struct xfs_mount	*mp = pip->i_mount;
 	struct xfs_mru_cache_elem *mru;
-	xfs_mount_t	*mp;
-	fstrm_item_t	*item;
-	xfs_agnumber_t	ag, rotorstep, startag;
-	int		err = 0;
+	xfs_agnumber_t		startag, ag;
 
 	ASSERT(S_ISDIR(pip->i_d.di_mode));
-	ASSERT(S_ISREG(ip->i_d.di_mode));
-	if (!S_ISDIR(pip->i_d.di_mode) || !S_ISREG(ip->i_d.di_mode))
-		return -EINVAL;
-
-	mp = pip->i_mount;
 
 	/*
-	 * We have a problem, Houston.
-	 *
-	 * Taking the iolock here violates inode locking order - we already
-	 * hold the ilock. Hence if we block getting this lock we may never
-	 * wake. Unfortunately, that means if we can't get the lock, we're
-	 * screwed in terms of getting a stream association - we can't spin
-	 * waiting for the lock because someone else is waiting on the lock we
-	 * hold and we cannot drop that as we are in a transaction here.
-	 *
-	 * Lucky for us, this inversion is not a problem because it's a
-	 * directory inode that we are trying to lock here.
-	 *
-	 * So, if we can't get the iolock without sleeping then just give up
+	 * If the directory already has a file stream associated we're done.
 	 */
-	if (!xfs_ilock_nowait(pip, XFS_IOLOCK_EXCL))
-		return 1;
-
-	/* If the parent directory is already in the cache, use its AG. */
 	mru = xfs_mru_cache_lookup(mp->m_filestream, pip->i_ino);
 	if (mru) {
-		item = container_of(mru, fstrm_item_t, mru);
-
-		ASSERT(item->ip == pip);
-		ag = item->ag;
 		xfs_mru_cache_done(mp->m_filestream);
-
-		TRACE_LOOKUP(mp, pip, pip, ag, xfs_filestream_peek_ag(mp, ag));
-		err = _xfs_filestream_update_ag(ip, pip, ag);
-
-		goto exit;
+		return 0;
 	}
 
 	/*
@@ -623,201 +387,107 @@ xfs_filestream_associate(
 	 * use the directory inode's AG.
 	 */
 	if (mp->m_flags & XFS_MOUNT_32BITINODES) {
-		rotorstep = xfs_rotorstep;
+		xfs_agnumber_t	 rotorstep = xfs_rotorstep;
 		startag = (mp->m_agfrotor / rotorstep) % mp->m_sb.sb_agcount;
 		mp->m_agfrotor = (mp->m_agfrotor + 1) %
 		                 (mp->m_sb.sb_agcount * rotorstep);
 	} else
 		startag = XFS_INO_TO_AGNO(mp, pip->i_ino);
 
-	/* Pick a new AG for the parent inode starting at startag. */
-	err = _xfs_filestream_pick_ag(mp, startag, &ag, 0, 0);
-	if (err || ag == NULLAGNUMBER)
-		goto exit_did_pick;
-
-	/* Associate the parent inode with the AG. */
-	err = _xfs_filestream_update_ag(pip, NULL, ag);
-	if (err)
-		goto exit_did_pick;
-
-	/* Associate the file inode with the AG. */
-	err = _xfs_filestream_update_ag(ip, pip, ag);
-	if (err)
-		goto exit_did_pick;
-
-	TRACE_ASSOCIATE(mp, ip, pip, ag, xfs_filestream_peek_ag(mp, ag));
-
-exit_did_pick:
-	/*
-	 * If _xfs_filestream_pick_ag() returned a valid AG, remove the
-	 * reference it took on it, since the file and directory will have taken
-	 * their own now if they were successfully cached.
-	 */
-	if (ag != NULLAGNUMBER)
-		xfs_filestream_put_ag(mp, ag);
-
-exit:
-	xfs_iunlock(pip, XFS_IOLOCK_EXCL);
-	return -err;
+	return xfs_filestream_pick_ag(pip, startag, &ag, 0, 0);
 }
 
 /*
- * Pick a new allocation group for the current file and its file stream.  This
- * function is called by xfs_bmap_filestreams() with the mount point's per-ag
- * lock held.
+ * Pick a new allocation group for the current file and its file stream.
+ *
+ * This is called when the allocator can't find a suitable extent in the
+ * current AG, and we have to move the stream into a new AG with more space.
  */
 int
 xfs_filestream_new_ag(
 	struct xfs_bmalloca	*ap,
 	xfs_agnumber_t		*agp)
 {
-	struct xfs_mru_cache_elem *mru, *mru2;
-	int		flags, err;
-	xfs_inode_t	*ip, *pip = NULL;
-	xfs_mount_t	*mp;
-	xfs_extlen_t	minlen;
-	fstrm_item_t	*dir, *file;
-	xfs_agnumber_t	ag = NULLAGNUMBER;
-
-	ip = ap->ip;
-	mp = ip->i_mount;
-	minlen = ap->length;
-	*agp = NULLAGNUMBER;
-
-	/*
-	 * Look for the file in the cache, removing it if it's found.  Doing
-	 * this allows it to be held across the dir lookup that follows.
-	 */
-	mru = xfs_mru_cache_remove(mp->m_filestream, ip->i_ino);
-	if (mru) {
-		file = container_of(mru, fstrm_item_t, mru);
-		ASSERT(ip == file->ip);
-
-		/* Save the file's parent inode and old AG number for later. */
-		pip = file->pip;
-		ag = file->ag;
-
-		/* Look for the file's directory in the cache. */
-		mru2 = xfs_mru_cache_lookup(mp->m_filestream, pip->i_ino);
-		if (mru2) {
-			dir = container_of(mru2, fstrm_item_t, mru);
-			ASSERT(pip == dir->ip);
-
-			/*
-			 * If the directory has already moved on to a new AG,
-			 * use that AG as the new AG for the file. Don't
-			 * forget to twiddle the AG refcounts to match the
-			 * movement.
-			 */
-			if (dir->ag != file->ag) {
-				xfs_filestream_put_ag(mp, file->ag);
-				xfs_filestream_get_ag(mp, dir->ag);
-				*agp = file->ag = dir->ag;
-			}
+	struct xfs_inode	*ip = ap->ip, *pip;
+	struct xfs_mount	*mp = ip->i_mount;
+	xfs_extlen_t		minlen = ap->length;
+	xfs_agnumber_t		startag = 0;
+	int			flags, err = 0;
+	struct xfs_mru_cache_elem *mru;
 
-			xfs_mru_cache_done(mp->m_filestream);
-		}
+	*agp = NULLAGNUMBER;
 
-		/*
-		 * Put the file back in the cache.  If this fails, the free
-		 * function needs to be called to tidy up in the same way as if
-		 * the item had simply expired from the cache.
-		 */
-		err = xfs_mru_cache_insert(mp->m_filestream, ip->i_ino, mru);
-		if (err) {
-			xfs_fstrm_free_func(mru);
-			return err;
-		}
+	pip = xfs_filestream_get_parent(ip);
+	if (!pip)
+		goto exit;
 
-		/*
-		 * If the file's AG was moved to the directory's new AG, there's
-		 * nothing more to be done.
-		 */
-		if (*agp != NULLAGNUMBER) {
-			TRACE_MOVEAG(mp, ip, pip,
-					ag, xfs_filestream_peek_ag(mp, ag),
-					*agp, xfs_filestream_peek_ag(mp, *agp));
-			return 0;
-		}
+	mru = xfs_mru_cache_remove(mp->m_filestream, pip->i_ino);
+	if (mru) {
+		struct xfs_fstrm_item *item =
+			container_of(mru, struct xfs_fstrm_item, mru);
+		startag = (item->ag + 1) % mp->m_sb.sb_agcount;
 	}
 
-	/*
-	 * If the file's parent directory is known, take its iolock in exclusive
-	 * mode to prevent two sibling files from racing each other to migrate
-	 * themselves and their parent to different AGs.
-	 *
-	 * Note that we lock the parent directory iolock inside the child
-	 * iolock here.  That's fine as we never hold both parent and child
-	 * iolock in any other place.  This is different from the ilock,
-	 * which requires locking of the child after the parent for namespace
-	 * operations.
-	 */
-	if (pip)
-		xfs_ilock(pip, XFS_IOLOCK_EXCL | XFS_IOLOCK_PARENT);
-
-	/*
-	 * A new AG needs to be found for the file.  If the file's parent
-	 * directory is also known, it will be moved to the new AG as well to
-	 * ensure that files created inside it in future use the new AG.
-	 */
-	ag = (ag == NULLAGNUMBER) ? 0 : (ag + 1) % mp->m_sb.sb_agcount;
 	flags = (ap->userdata ? XFS_PICK_USERDATA : 0) |
 	        (ap->flist->xbf_low ? XFS_PICK_LOWSPACE : 0);
 
-	err = _xfs_filestream_pick_ag(mp, ag, agp, flags, minlen);
-	if (err || *agp == NULLAGNUMBER)
-		goto exit;
+	err = xfs_filestream_pick_ag(pip, startag, agp, flags, minlen);
 
 	/*
-	 * If the file wasn't found in the file cache, then its parent directory
-	 * inode isn't known.  For this to have happened, the file must either
-	 * be pre-existing, or it was created long enough ago that its cache
-	 * entry has expired.  This isn't the sort of usage that the filestreams
-	 * allocator is trying to optimise, so there's no point trying to track
-	 * its new AG somehow in the filestream data structures.
+	 * Only free the item here so we skip over the old AG earlier.
 	 */
-	if (!pip) {
-		TRACE_ORPHAN(mp, ip, *agp);
-		goto exit;
-	}
+	if (mru)
+		xfs_fstrm_free_func(mru);
 
-	/* Associate the parent inode with the AG. */
-	err = _xfs_filestream_update_ag(pip, NULL, *agp);
-	if (err)
-		goto exit;
-
-	/* Associate the file inode with the AG. */
-	err = _xfs_filestream_update_ag(ip, pip, *agp);
-	if (err)
-		goto exit;
+	IRELE(pip);
+exit:
+	if (*agp == NULLAGNUMBER)
+		*agp = 0;
+	return err;
+}
 
-	TRACE_MOVEAG(mp, ip, pip, NULLAGNUMBER, 0,
-			*agp, xfs_filestream_peek_ag(mp, *agp));
+void
+xfs_filestream_deassociate(
+	struct xfs_inode	*ip)
+{
+	xfs_mru_cache_delete(ip->i_mount->m_filestream, ip->i_ino);
+}
 
-exit:
+int
+xfs_filestream_mount(
+	xfs_mount_t	*mp)
+{
 	/*
-	 * If _xfs_filestream_pick_ag() returned a valid AG, remove the
-	 * reference it took on it, since the file and directory will have taken
-	 * their own now if they were successfully cached.
+	 * The filestream timer tunable is currently fixed within the range of
+	 * one second to four minutes, with five seconds being the default.  The
+	 * group count is somewhat arbitrary, but it'd be nice to adhere to the
+	 * timer tunable to within about 10 percent.  This requires at least 10
+	 * groups.
 	 */
-	if (*agp != NULLAGNUMBER)
-		xfs_filestream_put_ag(mp, *agp);
-	else
-		*agp = 0;
+	return xfs_mru_cache_create(&mp->m_filestream, xfs_fstrm_centisecs * 10,
+				    10, xfs_fstrm_free_func);
+}
 
-	if (pip)
-		xfs_iunlock(pip, XFS_IOLOCK_EXCL);
+void
+xfs_filestream_unmount(
+	xfs_mount_t	*mp)
+{
+	xfs_mru_cache_destroy(mp->m_filestream);
+}
 
-	return err;
+
+/* needs to return a positive errno for the init path */
+int
+xfs_filestream_init(void)
+{
+	item_zone = kmem_zone_init(sizeof(struct xfs_fstrm_item), "fstrm_item");
+	if (!item_zone)
+		return -ENOMEM;
+	return 0;
 }
 
-/*
- * Remove an association between an inode and a filestream object.
- * Typically this is done on last close of an unlinked file.
- */
 void
-xfs_filestream_deassociate(
-	xfs_inode_t	*ip)
+xfs_filestream_uninit(void)
 {
-	xfs_mru_cache_delete(ip->i_mount->m_filestream, ip->i_ino);
+	kmem_zone_destroy(item_zone);
 }