JFS: Avoid parallel allocations within the same allocation group

When large files are writting in parallel, allocating the space for
these files within the same allocation group can cause severe
fragmentation of the files.  By keeping track of open, growing files
within an allocation group, we can force other new allocations into
a different allocation group to avoid this.

fs/jfs/file.c

@@ -19,6 +19,7 @@
 
 #include <linux/fs.h>
 #include "jfs_incore.h"
+#include "jfs_dmap.h"
 #include "jfs_txnmgr.h"
 #include "jfs_xattr.h"
 #include "jfs_debug.h"
@@ -94,6 +95,42 @@ static void jfs_truncate(struct inode *ip)
 	IWRITE_UNLOCK(ip);
 }
 
+static int jfs_open(struct inode *inode, struct file *file)
+{
+	int rc;
+
+	if ((rc = generic_file_open(inode, file)))
+		return rc;
+
+	/*
+	 * We attempt to allow only one "active" file open per aggregate
+	 * group.  Otherwise, appending to files in parallel can cause
+	 * fragmentation within the files.
+	 */
+	if (S_ISREG(inode->i_mode) && file->f_mode & FMODE_WRITE) {
+		struct jfs_inode_info *ji = JFS_IP(inode);
+		if (ji->active_ag == -1) {
+			ji->active_ag = ji->agno;
+			atomic_inc(
+			    &JFS_SBI(inode->i_sb)->bmap->db_active[ji->agno]);
+		}
+	}
+
+	return 0;
+}
+static int jfs_release(struct inode *inode, struct file *file)
+{
+	struct jfs_inode_info *ji = JFS_IP(inode);
+
+	if (ji->active_ag != -1) {
+		struct bmap *bmap = JFS_SBI(inode->i_sb)->bmap;
+		atomic_dec(&bmap->db_active[ji->active_ag]);
+		ji->active_ag = -1;
+	}
+
+	return 0;
+}
+
 struct inode_operations jfs_file_inode_operations = {
 	.truncate	= jfs_truncate,
 	.setxattr	= jfs_setxattr,
@@ -103,7 +140,7 @@ struct inode_operations jfs_file_inode_operations = {
 };
 
 struct file_operations jfs_file_operations = {
-	.open		= generic_file_open,
+	.open		= jfs_open,
 	.llseek		= generic_file_llseek,
 	.write		= generic_file_write,
 	.read		= generic_file_read,
@@ -112,4 +149,5 @@ struct file_operations jfs_file_operations = {
 	.writev		= generic_file_writev,
  	.sendfile	= generic_file_sendfile,
 	.fsync		= jfs_fsync,
+	.release	= jfs_release,
 };

fs/jfs/jfs_dmap.h

@@ -227,6 +227,7 @@ struct bmap {
 	struct dbmap db_bmap;	/* on-disk aggregate map descriptor */
 	struct inode *db_ipbmap;	/* ptr to aggregate map incore inode */
 	struct semaphore db_bmaplock;	/* aggregate map lock */
+	atomic_t db_active[MAXAG];	/* count of active, open files in AG */
 	u32 *db_DBmap;
 };
 

fs/jfs/jfs_extent.c

@@ -514,9 +514,12 @@ int extFill(struct inode *ip, xad_t * xp)
 static int
 extBalloc(struct inode *ip, s64 hint, s64 * nblocks, s64 * blkno)
 {
+	struct jfs_inode_info *ji = JFS_IP(ip);
+	struct jfs_sb_info *sbi = JFS_SBI(ip->i_sb);
 	s64 nb, nblks, daddr, max;
-	int rc, nbperpage = JFS_SBI(ip->i_sb)->nbperpage;
-	struct bmap *mp = JFS_SBI(ip->i_sb)->bmap;
+	int rc, nbperpage = sbi->nbperpage;
+	struct bmap *bmp = sbi->bmap;
+	int ag;
 
 	/* get the number of blocks to initially attempt to allocate.
 	 * we'll first try the number of blocks requested unless this
@@ -524,7 +527,7 @@ extBalloc(struct inode *ip, s64 hint, s64 * nblocks, s64 * blkno)
 	 * blocks in the map. in that case, we'll start off with the 
 	 * maximum free.
 	 */
-	max = (s64) 1 << mp->db_maxfreebud;
+	max = (s64) 1 << bmp->db_maxfreebud;
 	if (*nblocks >= max && *nblocks > nbperpage)
 		nb = nblks = (max > nbperpage) ? max : nbperpage;
 	else
@@ -549,6 +552,18 @@ extBalloc(struct inode *ip, s64 hint, s64 * nblocks, s64 * blkno)
 	*nblocks = nb;
 	*blkno = daddr;
 
+	if (S_ISREG(ip->i_mode) && (ji->fileset == FILESYSTEM_I)) {
+		ag = BLKTOAG(daddr, sbi);
+		if (ji->active_ag == -1) {
+			atomic_inc(&bmp->db_active[ag]);
+			ji->active_ag = ag;
+		} else if (ji->active_ag != ag) {
+			atomic_dec(&bmp->db_active[ji->active_ag]);
+			atomic_inc(&bmp->db_active[ag]);
+			ji->active_ag = ag;
+		}
+	}
+
 	return (0);
 }
 

fs/jfs/jfs_imap.c

@@ -429,6 +429,7 @@ int diRead(struct inode *ip)
 
 	/* set the ag for the inode */
 	JFS_IP(ip)->agno = BLKTOAG(agstart, sbi);
+	JFS_IP(ip)->active_ag = -1;
 
 	return (rc);
 }
@@ -1358,6 +1359,7 @@ diInitInode(struct inode *ip, int iagno, int ino, int extno, struct iag * iagp)
 	DBG_DIALLOC(JFS_IP(ipimap)->i_imap, ip->i_ino);
 	jfs_ip->ixpxd = iagp->inoext[extno];
 	jfs_ip->agno = BLKTOAG(le64_to_cpu(iagp->agstart), sbi);
+	jfs_ip->active_ag = -1;
 }
 
 
@@ -1413,6 +1415,21 @@ int diAlloc(struct inode *pip, boolean_t dir, struct inode *ip)
 	 * moving backward on the disk.)  compute the hint within the
 	 * file system and the iag.
 	 */
+
+	/* get the ag number of this iag */
+	agno = JFS_IP(pip)->agno;
+
+	if (atomic_read(&JFS_SBI(pip->i_sb)->bmap->db_active[agno])) {
+		/*
+		 * There is an open file actively growing.  We want to
+		 * allocate new inodes from a different ag to avoid
+		 * fragmentation problems.
+		 */
+		agno = dbNextAG(JFS_SBI(pip->i_sb)->ipbmap);
+		AG_LOCK(imap, agno);
+		goto tryag;
+	}
+
 	inum = pip->i_ino + 1;
 	ino = inum & (INOSPERIAG - 1);
 
@@ -1420,9 +1437,6 @@ int diAlloc(struct inode *pip, boolean_t dir, struct inode *ip)
 	if (ino == 0)
 		inum = pip->i_ino;
 
-	/* get the ag number of this iag */
-	agno = JFS_IP(pip)->agno;
-
 	/* lock the AG inode map information */
 	AG_LOCK(imap, agno);
 

fs/jfs/jfs_incore.h

@@ -49,7 +49,7 @@ struct jfs_inode_info {
 	long	cflag;		/* commit flags		*/
 	u16	bxflag;		/* xflag of pseudo buffer?	*/
 	unchar	agno;		/* ag number			*/
-	unchar	pad;		/* pad			*/
+	signed char active_ag;	/* ag currently allocating from	*/
 	lid_t	blid;		/* lid of pseudo buffer?	*/
 	lid_t	atlhead;	/* anonymous tlock list head	*/
 	lid_t	atltail;	/* anonymous tlock list tail	*/

fs/jfs/super.c

@@ -406,6 +406,7 @@ static void init_once(void *foo, kmem_cache_t * cachep, unsigned long flags)
 		init_rwsem(&jfs_ip->rdwrlock);
 		init_MUTEX(&jfs_ip->commit_sem);
 		jfs_ip->atlhead = 0;
+		jfs_ip->active_ag = -1;
 		inode_init_once(&jfs_ip->vfs_inode);
 	}
 }