[XFS] Concurrent Multi-File Data Streams

In media spaces, video is often stored in a frame-per-file format. When
dealing with uncompressed realtime HD video streams in this format, it is
crucial that files do not get fragmented and that multiple files a placed
contiguously on disk.

When multiple streams are being ingested and played out at the same time,
it is critical that the filesystem does not cross the streams and
interleave them together as this creates seek and readahead cache miss
latency and prevents both ingest and playout from meeting frame rate
targets.

This patch set creates a "stream of files" concept into the allocator to
place all the data from a single stream contiguously on disk so that RAID
array readahead can be used effectively. Each additional stream gets
placed in different allocation groups within the filesystem, thereby
ensuring that we don't cross any streams. When an AG fills up, we select a
new AG for the stream that is not in use.

The core of the functionality is the stream tracking - each inode that we
create in a directory needs to be associated with the directories' stream.
Hence every time we create a file, we look up the directories' stream
object and associate the new file with that object.

Once we have a stream object for a file, we use the AG that the stream
object point to for allocations. If we can't allocate in that AG (e.g. it
is full) we move the entire stream to another AG. Other inodes in the same
stream are moved to the new AG on their next allocation (i.e. lazy
update).

Stream objects are kept in a cache and hold a reference on the inode.
Hence the inode cannot be reclaimed while there is an outstanding stream
reference. This means that on unlink we need to remove the stream
association and we also need to flush all the associations on certain
events that want to reclaim all unreferenced inodes (e.g. filesystem
freeze).

SGI-PV: 964469
SGI-Modid: xfs-linux-melb:xfs-kern:29096a

Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Barry Naujok <bnaujok@sgi.com>
Signed-off-by: Donald Douwsma <donaldd@sgi.com>
Signed-off-by: Christoph Hellwig <hch@infradead.org>
Signed-off-by: Tim Shimmin <tes@sgi.com>
Signed-off-by: Vlad Apostolov <vapo@sgi.com>

21 files changed, 1730 insertions, 12 deletions

diff --git a/fs/xfs/Makefile-linux-2.6 b/fs/xfs/Makefile-linux-2.6
index b49989bb89a..e7a9a83f008 100644
--- a/fs/xfs/Makefile-linux-2.6
+++ b/fs/xfs/Makefile-linux-2.6
@@ -64,6 +64,7 @@ xfs-y				+= xfs_alloc.o \
 				   xfs_dir2_sf.o \
 				   xfs_error.o \
 				   xfs_extfree_item.o \
+				   xfs_filestream.o \
 				   xfs_fsops.o \
 				   xfs_ialloc.o \
 				   xfs_ialloc_btree.o \
@@ -77,6 +78,7 @@ xfs-y				+= xfs_alloc.o \
 				   xfs_log.o \
 				   xfs_log_recover.o \
 				   xfs_mount.o \
+				   xfs_mru_cache.o \
 				   xfs_rename.o \
 				   xfs_trans.o \
 				   xfs_trans_ail.o \
diff --git a/fs/xfs/linux-2.6/xfs_globals.c b/fs/xfs/linux-2.6/xfs_globals.c
index ed3a5e1b4b6..bb72c3d4141 100644
--- a/fs/xfs/linux-2.6/xfs_globals.c
+++ b/fs/xfs/linux-2.6/xfs_globals.c
@@ -46,6 +46,7 @@ xfs_param_t xfs_params = {
 	.inherit_nosym	= {	0,		0,		1	},
 	.rotorstep	= {	1,		1,		255	},
 	.inherit_nodfrg	= {	0,		1,		1	},
+	.fstrm_timer	= {	1,		50,		3600*100},
 };
 
 /*
diff --git a/fs/xfs/linux-2.6/xfs_linux.h b/fs/xfs/linux-2.6/xfs_linux.h
index af24a457d3a..330c4ba9d40 100644
--- a/fs/xfs/linux-2.6/xfs_linux.h
+++ b/fs/xfs/linux-2.6/xfs_linux.h
@@ -123,6 +123,7 @@
 #define xfs_inherit_nosymlinks	xfs_params.inherit_nosym.val
 #define xfs_rotorstep		xfs_params.rotorstep.val
 #define xfs_inherit_nodefrag	xfs_params.inherit_nodfrg.val
+#define xfs_fstrm_centisecs	xfs_params.fstrm_timer.val
 
 #define current_cpu()		(raw_smp_processor_id())
 #define current_pid()		(current->pid)
diff --git a/fs/xfs/linux-2.6/xfs_sysctl.c b/fs/xfs/linux-2.6/xfs_sysctl.c
index cd6eaa44aa2..bb997d75c05 100644
--- a/fs/xfs/linux-2.6/xfs_sysctl.c
+++ b/fs/xfs/linux-2.6/xfs_sysctl.c
@@ -210,6 +210,17 @@ static ctl_table xfs_table[] = {
 		.extra1		= &xfs_params.inherit_nodfrg.min,
 		.extra2		= &xfs_params.inherit_nodfrg.max
 	},
+	{
+		.ctl_name	= XFS_FILESTREAM_TIMER,
+		.procname	= "filestream_centisecs",
+		.data		= &xfs_params.fstrm_timer.val,
+		.maxlen		= sizeof(int),
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec_minmax,
+		.strategy	= &sysctl_intvec,
+		.extra1		= &xfs_params.fstrm_timer.min,
+		.extra2		= &xfs_params.fstrm_timer.max,
+	},
 	/* please keep this the last entry */
 #ifdef CONFIG_PROC_FS
 	{
diff --git a/fs/xfs/linux-2.6/xfs_sysctl.h b/fs/xfs/linux-2.6/xfs_sysctl.h
index a631fb8cc5a..98b97e399d6 100644
--- a/fs/xfs/linux-2.6/xfs_sysctl.h
+++ b/fs/xfs/linux-2.6/xfs_sysctl.h
@@ -47,6 +47,7 @@ typedef struct xfs_param {
 	xfs_sysctl_val_t inherit_nosym;	/* Inherit the "nosymlinks" flag. */
 	xfs_sysctl_val_t rotorstep;	/* inode32 AG rotoring control knob */
 	xfs_sysctl_val_t inherit_nodfrg;/* Inherit the "nodefrag" inode flag. */
+	xfs_sysctl_val_t fstrm_timer;	/* Filestream dir-AG assoc'n timeout. */
 } xfs_param_t;
 
 /*
@@ -86,6 +87,7 @@ enum {
 	XFS_INHERIT_NOSYM = 19,
 	XFS_ROTORSTEP = 20,
 	XFS_INHERIT_NODFRG = 21,
+	XFS_FILESTREAM_TIMER = 22,
 };
 
 extern xfs_param_t	xfs_params;
diff --git a/fs/xfs/xfs.h b/fs/xfs/xfs.h
index bf0a12040b1..b5a7d92c684 100644
--- a/fs/xfs/xfs.h
+++ b/fs/xfs/xfs.h
@@ -38,6 +38,7 @@
 #define XFS_RW_TRACE 1
 #define XFS_BUF_TRACE 1
 #define XFS_VNODE_TRACE 1
+#define XFS_FILESTREAMS_TRACE 1
 #endif
 
 #include <linux-2.6/xfs_linux.h>
diff --git a/fs/xfs/xfs_ag.h b/fs/xfs/xfs_ag.h
index b1dd0029c60..51c09c114a2 100644
--- a/fs/xfs/xfs_ag.h
+++ b/fs/xfs/xfs_ag.h
@@ -196,6 +196,7 @@ typedef struct xfs_perag
 	lock_t		pagb_lock;	/* lock for pagb_list */
 #endif
 	xfs_perag_busy_t *pagb_list;	/* unstable blocks */
+	atomic_t        pagf_fstrms;    /* # of filestreams active in this AG */
 } xfs_perag_t;
 
 #define	XFS_AG_MAXLEVELS(mp)		((mp)->m_ag_maxlevels)
diff --git a/fs/xfs/xfs_bmap.c b/fs/xfs/xfs_bmap.c
index 09d86388bb7..51ba689a455 100644
--- a/fs/xfs/xfs_bmap.c
+++ b/fs/xfs/xfs_bmap.c
@@ -52,6 +52,7 @@
 #include "xfs_quota.h"
 #include "xfs_trans_space.h"
 #include "xfs_buf_item.h"
+#include "xfs_filestream.h"
 
 
 #ifdef DEBUG
@@ -2725,9 +2726,15 @@ xfs_bmap_btalloc(
 	}
 	nullfb = ap->firstblock == NULLFSBLOCK;
 	fb_agno = nullfb ? NULLAGNUMBER : XFS_FSB_TO_AGNO(mp, ap->firstblock);
-	if (nullfb)
-		ap->rval = XFS_INO_TO_FSB(mp, ap->ip->i_ino);
-	else
+	if (nullfb) {
+		if (ap->userdata && xfs_inode_is_filestream(ap->ip)) {
+			ag = xfs_filestream_lookup_ag(ap->ip);
+			ag = (ag != NULLAGNUMBER) ? ag : 0;
+			ap->rval = XFS_AGB_TO_FSB(mp, ag, 0);
+		} else {
+			ap->rval = XFS_INO_TO_FSB(mp, ap->ip->i_ino);
+		}
+	} else
 		ap->rval = ap->firstblock;
 
 	xfs_bmap_adjacent(ap);
@@ -2751,13 +2758,22 @@ xfs_bmap_btalloc(
 	args.firstblock = ap->firstblock;
 	blen = 0;
 	if (nullfb) {
-		args.type = XFS_ALLOCTYPE_START_BNO;
+		if (ap->userdata && xfs_inode_is_filestream(ap->ip))
+			args.type = XFS_ALLOCTYPE_NEAR_BNO;
+		else
+			args.type = XFS_ALLOCTYPE_START_BNO;
 		args.total = ap->total;
+
 		/*
-		 * Find the longest available space.
-		 * We're going to try for the whole allocation at once.
+		 * Search for an allocation group with a single extent
+		 * large enough for the request.
+		 *
+		 * If one isn't found, then adjust the minimum allocation
+		 * size to the largest space found.
 		 */
 		startag = ag = XFS_FSB_TO_AGNO(mp, args.fsbno);
+		if (startag == NULLAGNUMBER)
+			startag = ag = 0;
 		notinit = 0;
 		down_read(&mp->m_peraglock);
 		while (blen < ap->alen) {
@@ -2783,6 +2799,35 @@ xfs_bmap_btalloc(
 					blen = longest;
 			} else
 				notinit = 1;
+
+			if (xfs_inode_is_filestream(ap->ip)) {
+				if (blen >= ap->alen)
+					break;
+
+				if (ap->userdata) {
+					/*
+					 * If startag is an invalid AG, we've
+					 * come here once before and
+					 * xfs_filestream_new_ag picked the
+					 * best currently available.
+					 *
+					 * Don't continue looping, since we
+					 * could loop forever.
+					 */
+					if (startag == NULLAGNUMBER)
+						break;
+
+					error = xfs_filestream_new_ag(ap, &ag);
+					if (error) {
+						up_read(&mp->m_peraglock);
+						return error;
+					}
+
+					/* loop again to set 'blen'*/
+					startag = NULLAGNUMBER;
+					continue;
+				}
+			}
 			if (++ag == mp->m_sb.sb_agcount)
 				ag = 0;
 			if (ag == startag)
@@ -2807,8 +2852,18 @@ xfs_bmap_btalloc(
 		 */
 		else
 			args.minlen = ap->alen;
+
+		/*
+		 * set the failure fallback case to look in the selected
+		 * AG as the stream may have moved.
+		 */
+		if (xfs_inode_is_filestream(ap->ip))
+			ap->rval = args.fsbno = XFS_AGB_TO_FSB(mp, ag, 0);
 	} else if (ap->low) {
-		args.type = XFS_ALLOCTYPE_START_BNO;
+		if (xfs_inode_is_filestream(ap->ip))
+			args.type = XFS_ALLOCTYPE_FIRST_AG;
+		else
+			args.type = XFS_ALLOCTYPE_START_BNO;
 		args.total = args.minlen = ap->minlen;
 	} else {
 		args.type = XFS_ALLOCTYPE_NEAR_BNO;
diff --git a/fs/xfs/xfs_clnt.h b/fs/xfs/xfs_clnt.h
index 5b7eb81453b..f89196cb08d 100644
--- a/fs/xfs/xfs_clnt.h
+++ b/fs/xfs/xfs_clnt.h
@@ -99,5 +99,7 @@ struct xfs_mount_args {
  */
 #define XFSMNT2_COMPAT_IOSIZE	0x00000001	/* don't report large preferred
 						 * I/O size in stat(2) */
+#define XFSMNT2_FILESTREAMS	0x00000002	/* enable the filestreams
+						 * allocator */
 
 #endif	/* __XFS_CLNT_H__ */
diff --git a/fs/xfs/xfs_dinode.h b/fs/xfs/xfs_dinode.h
index b33826961c4..fefd0116bac 100644
--- a/fs/xfs/xfs_dinode.h
+++ b/fs/xfs/xfs_dinode.h
@@ -257,6 +257,7 @@ typedef enum xfs_dinode_fmt
 #define XFS_DIFLAG_EXTSIZE_BIT      11	/* inode extent size allocator hint */
 #define XFS_DIFLAG_EXTSZINHERIT_BIT 12	/* inherit inode extent size */
 #define XFS_DIFLAG_NODEFRAG_BIT     13	/* do not reorganize/defragment */
+#define XFS_DIFLAG_FILESTREAM_BIT   14  /* use filestream allocator */
 #define XFS_DIFLAG_REALTIME      (1 << XFS_DIFLAG_REALTIME_BIT)
 #define XFS_DIFLAG_PREALLOC      (1 << XFS_DIFLAG_PREALLOC_BIT)
 #define XFS_DIFLAG_NEWRTBM       (1 << XFS_DIFLAG_NEWRTBM_BIT)
@@ -271,12 +272,13 @@ typedef enum xfs_dinode_fmt
 #define XFS_DIFLAG_EXTSIZE       (1 << XFS_DIFLAG_EXTSIZE_BIT)
 #define XFS_DIFLAG_EXTSZINHERIT  (1 << XFS_DIFLAG_EXTSZINHERIT_BIT)
 #define XFS_DIFLAG_NODEFRAG      (1 << XFS_DIFLAG_NODEFRAG_BIT)
+#define XFS_DIFLAG_FILESTREAM    (1 << XFS_DIFLAG_FILESTREAM_BIT)
 
 #define XFS_DIFLAG_ANY \
 	(XFS_DIFLAG_REALTIME | XFS_DIFLAG_PREALLOC | XFS_DIFLAG_NEWRTBM | \
 	 XFS_DIFLAG_IMMUTABLE | XFS_DIFLAG_APPEND | XFS_DIFLAG_SYNC | \
 	 XFS_DIFLAG_NOATIME | XFS_DIFLAG_NODUMP | XFS_DIFLAG_RTINHERIT | \
 	 XFS_DIFLAG_PROJINHERIT | XFS_DIFLAG_NOSYMLINKS | XFS_DIFLAG_EXTSIZE | \
-	 XFS_DIFLAG_EXTSZINHERIT | XFS_DIFLAG_NODEFRAG)
+	 XFS_DIFLAG_EXTSZINHERIT | XFS_DIFLAG_NODEFRAG | XFS_DIFLAG_FILESTREAM)
 
 #endif	/* __XFS_DINODE_H__ */
diff --git a/fs/xfs/xfs_filestream.c b/fs/xfs/xfs_filestream.c
new file mode 100644
index 00000000000..ce2278611bb
--- /dev/null
+++ b/fs/xfs/xfs_filestream.c
@@ -0,0 +1,771 @@
+/*
+ * Copyright (c) 2006-2007 Silicon Graphics, Inc.
+ * All Rights Reserved.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it would be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write the Free Software Foundation,
+ * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
+ */
+#include "xfs.h"
+#include "xfs_bmap_btree.h"
+#include "xfs_inum.h"
+#include "xfs_dir2.h"
+#include "xfs_dir2_sf.h"
+#include "xfs_attr_sf.h"
+#include "xfs_dinode.h"
+#include "xfs_inode.h"
+#include "xfs_ag.h"
+#include "xfs_dmapi.h"
+#include "xfs_log.h"
+#include "xfs_trans.h"
+#include "xfs_sb.h"
+#include "xfs_mount.h"
+#include "xfs_bmap.h"
+#include "xfs_alloc.h"
+#include "xfs_utils.h"
+#include "xfs_mru_cache.h"
+#include "xfs_filestream.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, __FUNCTION__, __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
+{
+	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;
+
+
+/*
+ * 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)
+{
+	int		err, trylock, nscan;
+	xfs_extlen_t	delta, longest, need, free, minfree, maxfree = 0;
+	xfs_agnumber_t	ag, max_ag = NULLAGNUMBER;
+	struct xfs_perag *pag;
+
+	/* 2% of an AG's blocks must be free for it to be chosen. */
+	minfree = mp->m_sb.sb_agblocks / 50;
+
+	ag = startag;
+	*agp = NULLAGNUMBER;
+
+	/* For the first pass, don't sleep trying to init the per-AG. */
+	trylock = XFS_ALLOC_FLAG_TRYLOCK;
+
+	for (nscan = 0; 1; nscan++) {
+
+		TRACE_AG_SCAN(mp, ag, xfs_filestream_peek_ag(mp, ag));
+
+		pag = mp->m_perag + ag;
+
+		if (!pag->pagf_init) {
+			err = xfs_alloc_pagf_init(mp, NULL, ag, trylock);
+			if (err && !trylock)
+				return err;
+		}
+
+		/* Might fail sometimes during the 1st pass with trylock set. */
+		if (!pag->pagf_init)
+			goto next_ag;
+
+		/* Keep track of the AG with the most free blocks. */
+		if (pag->pagf_freeblks > maxfree) {
+			maxfree = pag->pagf_freeblks;
+			max_ag = ag;
+		}
+
+		/*
+		 * The AG reference count does two things: it enforces mutual
+		 * exclusion when examining the suitability of an AG in this
+		 * loop, and it guards against two filestreams being established
+		 * in the same AG as each other.
+		 */
+		if (xfs_filestream_get_ag(mp, ag) > 1) {
+			xfs_filestream_put_ag(mp, ag);
+			goto next_ag;
+		}
+
+		need = XFS_MIN_FREELIST_PAG(pag, mp);
+		delta = need > pag->pagf_flcount ? need - pag->pagf_flcount : 0;
+		longest = (pag->pagf_longest > delta) ?
+		          (pag->pagf_longest - delta) :
+		          (pag->pagf_flcount > 0 || pag->pagf_longest > 0);
+
+		if (((minlen && longest >= minlen) ||
+		     (!minlen && pag->pagf_freeblks >= minfree)) &&
+		    (!pag->pagf_metadata || !(flags & XFS_PICK_USERDATA) ||
+		     (flags & XFS_PICK_LOWSPACE))) {
+
+			/* Break out, retaining the reference on the AG. */
+			free = pag->pagf_freeblks;
+			*agp = ag;
+			break;
+		}
+
+		/* Drop the reference on this AG, it's not usable. */
+		xfs_filestream_put_ag(mp, ag);
+next_ag:
+		/* Move to the next AG, wrapping to AG 0 if necessary. */
+		if (++ag >= mp->m_sb.sb_agcount)
+			ag = 0;
+
+		/* If a full pass of the AGs hasn't been done yet, continue. */
+		if (ag != startag)
+			continue;
+
+		/* Allow sleeping in xfs_alloc_pagf_init() on the 2nd pass. */
+		if (trylock != 0) {
+			trylock = 0;
+			continue;
+		}
+
+		/* Finally, if lowspace wasn't set, set it for the 3rd pass. */
+		if (!(flags & XFS_PICK_LOWSPACE)) {
+			flags |= XFS_PICK_LOWSPACE;
+			continue;
+		}
+
+		/*
+		 * Take the AG with the most free space, regardless of whether
+		 * it's already in use by another filestream.
+		 */
+		if (max_ag != NULLAGNUMBER) {
+			xfs_filestream_get_ag(mp, max_ag);
+			TRACE_AG_PICK1(mp, max_ag, maxfree);
+			free = maxfree;
+			*agp = max_ag;
+			break;
+		}
+
+		/* take AG 0 if none matched */
+		TRACE_AG_PICK1(mp, max_ag, maxfree);
+		*agp = 0;
+		return 0;
+	}
+
+	TRACE_AG_PICK2(mp, startag, *agp, xfs_filestream_peek_ag(mp, *agp),
+			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.  Must be called with the
+ * m_peraglock held in read mode.
+ */
+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;
+	xfs_mru_cache_t	*cache;
+	fstrm_item_t	*item;
+	xfs_agnumber_t	old_ag;
+	xfs_inode_t	*old_pip;
+
+	/*
+	 * Either ip is a regular file and pip is a directory, or ip is a
+	 * directory and pip is NULL.
+	 */
+	ASSERT(ip && (((ip->i_d.di_mode & S_IFREG) && pip &&
+	               (pip->i_d.di_mode & S_IFDIR)) ||
+	              ((ip->i_d.di_mode & S_IFDIR) && !pip)));
+
+	mp = ip->i_mount;
+	cache = mp->m_filestream;
+
+	item = xfs_mru_cache_lookup(cache, ip->i_ino);
+	if (item) {
+		ASSERT(item->ip == ip);
+		old_ag = item->ag;
+		item->ag = ag;
+		old_pip = item->pip;
+		item->pip = pip;
+		xfs_mru_cache_done(cache);
+
+		/*
+		 * 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));
+		return 0;
+	}
+
+	item = kmem_zone_zalloc(item_zone, KM_MAYFAIL);
+	if (!item)
+		return ENOMEM;
+
+	item->ag = ag;
+	item->ip = ip;
+	item->pip = pip;
+
+	err = xfs_mru_cache_insert(cache, ip->i_ino, item);
+	if (err) {
+		kmem_zone_free(item_zone, item);
+		return err;
+	}
+
+	/* 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. */
+void
+xfs_fstrm_free_func(
+	xfs_ino_t	ino,
+	fstrm_item_t	*item)
+{
+	xfs_inode_t	*ip = item->ip;
+	int ref;
+
+	ASSERT(ip->i_ino == ino);
+
+	xfs_iflags_clear(ip, XFS_IFILESTREAM);
+
+	/* Drop the reference taken on the AG when the item was added. */
+	ref = xfs_filestream_put_ag(ip->i_mount, item->ag);
+
+	ASSERT(ref >= 0);
+	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. */
+	kmem_zone_free(item_zone, item);
+}
+
+/*
+ * 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");
+#ifdef XFS_FILESTREAMS_TRACE
+	xfs_filestreams_trace_buf = ktrace_alloc(XFS_FSTRM_KTRACE_SIZE, KM_SLEEP);
+#endif
+	return item_zone ? 0 : -ENOMEM;
+}
+
+/*
+ * 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)
+{
+#ifdef XFS_FILESTREAMS_TRACE
+	ktrace_free(xfs_filestreams_trace_buf);
+#endif
+	kmem_zone_destroy(item_zone);
+}
+
+/*
+ * 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;
+
+	/*
+	 * 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;
+
+	err = xfs_mru_cache_create(&mp->m_filestream, lifetime, grp_count,
+	                     (xfs_mru_cache_free_func_t)xfs_fstrm_free_func);
+
+	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);
+}
+
+/*
+ * If the mount point's m_perag array is going to be reallocated, all
+ * outstanding cache entries must be flushed to avoid accessing reference count
+ * addresses that have been freed.  The call to xfs_filestream_flush() must be
+ * made inside the block that holds the m_peraglock in write mode to do the
+ * reallocation.
+ */
+void
+xfs_filestream_flush(
+	xfs_mount_t	*mp)
+{
+	/* point in time flush, so keep the reaper running */
+	xfs_mru_cache_flush(mp->m_filestream, 1);
+}
+
+/*
+ * Return the AG of the filestream the file or directory belongs to, or
+ * NULLAGNUMBER otherwise.
+ */
+xfs_agnumber_t
+xfs_filestream_lookup_ag(
+	xfs_inode_t	*ip)
+{
+	xfs_mru_cache_t	*cache;
+	fstrm_item_t	*item;
+	xfs_agnumber_t	ag;
+	int		ref;
+
+	if (!(ip->i_d.di_mode & (S_IFREG | S_IFDIR))) {
+		ASSERT(0);
+		return NULLAGNUMBER;
+	}
+
+	cache = ip->i_mount->m_filestream;
+	item = xfs_mru_cache_lookup(cache, ip->i_ino);
+	if (!item) {
+		TRACE_LOOKUP(ip->i_mount, ip, NULL, NULLAGNUMBER, 0);
+		return NULLAGNUMBER;
+	}
+
+	ASSERT(ip == item->ip);
+	ag = item->ag;
+	ref = xfs_filestream_peek_ag(ip->i_mount, ag);
+	xfs_mru_cache_done(cache);
+
+	TRACE_LOOKUP(ip->i_mount, ip, item->pip, ag, ref);
+	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.
+ *
+ * Returns < 0 on error, 0 if successful association occurred, > 0 if
+ * we failed to get an association because of locking issues.
+ */
+int
+xfs_filestream_associate(
+	xfs_inode_t	*pip,
+	xfs_inode_t	*ip)
+{
+	xfs_mount_t	*mp;
+	xfs_mru_cache_t	*cache;
+	fstrm_item_t	*item;
+	xfs_agnumber_t	ag, rotorstep, startag;
+	int		err = 0;
+
+	ASSERT(pip->i_d.di_mode & S_IFDIR);
+	ASSERT(ip->i_d.di_mode & S_IFREG);
+	if (!(pip->i_d.di_mode & S_IFDIR) || !(ip->i_d.di_mode & S_IFREG))
+		return -EINVAL;
+
+	mp = pip->i_mount;
+	cache = mp->m_filestream;
+	down_read(&mp->m_peraglock);
+
+	/*
+	 * 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 rarely a problem because it's a
+	 * directory inode that we are trying to lock here and that means the
+	 * only place that matters is xfs_sync_inodes() and SYNC_DELWRI is
+	 * used. i.e. freeze, remount-ro, quotasync or unmount.
+	 *
+	 * So, if we can't get the iolock without sleeping then just give up
+	 */
+	if (!xfs_ilock_nowait(pip, XFS_IOLOCK_EXCL)) {
+		up_read(&mp->m_peraglock);
+		return 1;
+	}
+
+	/* If the parent directory is already in the cache, use its AG. */
+	item = xfs_mru_cache_lookup(cache, pip->i_ino);
+	if (item) {
+		ASSERT(item->ip == pip);
+		ag = item->ag;
+		xfs_mru_cache_done(cache);
+
+		TRACE_LOOKUP(mp, pip, pip, ag, xfs_filestream_peek_ag(mp, ag));
+		err = _xfs_filestream_update_ag(ip, pip, ag);
+
+		goto exit;
+	}
+
+	/*
+	 * Set the starting AG using the rotor for inode32, otherwise
+	 * use the directory inode's AG.
+	 */
+	if (mp->m_flags & XFS_MOUNT_32BITINODES) {
+		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);
+	up_read(&mp->m_peraglock);
+	return -err;
+}
+
+/*
+ * 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.
+ */
+int
+xfs_filestream_new_ag(