Skip to contentSkip to conversation

Improve inode list scanning in add_dquot_ref()

PATCH In review Jan Kara 2 participants 3 messages Dec 11, 2007 – Dec 12, 2007 VFS

lore
Share & tools
lore

Jan Kara

author reviewer
#0 lore raw

PATCH 1/1 Improve inode list scanning in add_dquot_ref()

Hi,

attached patch makes add_dquot_ref() quite faster if used on a life
filesystem. It has survived some testing on my machine but there could be some subtle races I've missed so please leave it in -mm for a while (not that I'd think that people use quotas with -mm kernels but at least it should go into vanilla in -rc1...). Thanks.

Honza

---
We restarted scan of sb->s_inodes list whenever we had to drop inode_lock in add_dquot_ref(). This leads to overall quadratic running time and thus add_dquot_ref() can take several minutes when called on a life filesystem. We fix the problem by using the fact that inode cannot be removed from s_inodes list while we hold a reference to it and thus we can safely restart the scan if we don't drop the reference. Here we use the fact that inodes freshly added to s_inodes list are already guaranteed to have quotas properly initialized and the ordering of inodes on s_inodes list does not change so we cannot skip any inode. Thanks goes to Nick <[email protected]> for analyzing the problem and testing the fix.

fs/dquot.c +12 -5
index 2809768..ecea4be 100644
@@ -696,9 +696,8 @@ static int dqinit_needed(struct inode *inode, int type)
 /* This routine is guarded by dqonoff_mutex mutex */
 static void add_dquot_ref(struct super_block *sb, int type)
 {
-	struct inode *inode;
+	struct inode *inode, *old_inode = NULL;
 
-restart:
 	spin_lock(&inode_lock);
 	list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
 		if (!atomic_read(&inode->i_writecount))
@@ -711,12 +710,20 @@ restart:
 		__iget(inode);
 		spin_unlock(&inode_lock);
 
+		if (old_inode)
+			iput(old_inode);
 		sb->dq_op->initialize(inode, type);
-		iput(inode);
-		/* As we may have blocked we had better restart... */
-		goto restart;
+		/* We hold a reference to 'inode' so it couldn't have been
+		 * removed from s_inodes list while we dropped the inode_lock.
+		 * We cannot iput the inode now as we can be holding the last
+		 * reference and we cannot iput it under inode_lock. So we
+		 * keep the reference and iput it later. */
+		old_inode = inode;
+		spin_lock(&inode_lock);
 	}
 	spin_unlock(&inode_lock);
+	if (old_inode)
+		iput(old_inode);
 }
 
 /* Return 0 if dqput() won't block (note that 1 doesn't necessarily mean blocking) */
#1 lore raw
Quoted 53 lines · On Tue, 11 Dec 2007 14:43:19 +0100 Jan Kara <[email protected]> wrote:

Hi,

attached patch makes add_dquot_ref() quite faster if used on a life
filesystem. It has survived some testing on my machine but there could be some subtle races I've missed so please leave it in -mm for a while (not that I'd think that people use quotas with -mm kernels but at least it should go into vanilla in -rc1...). Thanks.

Honza

---
We restarted scan of sb->s_inodes list whenever we had to drop inode_lock in add_dquot_ref(). This leads to overall quadratic running time and thus add_dquot_ref() can take several minutes when called on a life filesystem. We fix the problem by using the fact that inode cannot be removed from s_inodes list while we hold a reference to it and thus we can safely restart the scan if we don't drop the reference. Here we use the fact that inodes freshly added to s_inodes list are already guaranteed to have quotas properly initialized and the ordering of inodes on s_inodes list does not change so we cannot skip any inode. Thanks goes to Nick <[email protected]> for analyzing the problem and testing the fix.

fs/dquot.c +6 -5
index 2809768..ecea4be 100644
@@ -696,9 +696,8 @@ static int dqinit_needed(struct inode *inode, int type)
 /* This routine is guarded by dqonoff_mutex mutex */
 static void add_dquot_ref(struct super_block *sb, int type)
 {
-	struct inode *inode;
+	struct inode *inode, *old_inode = NULL;
 
-restart:
 	spin_lock(&inode_lock);
 	list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
 		if (!atomic_read(&inode->i_writecount))
@@ -711,12 +710,20 @@ restart:
 		__iget(inode);
 		spin_unlock(&inode_lock);
 
+		if (old_inode)
+			iput(old_inode);
 		sb->dq_op->initialize(inode, type);
-		iput(inode);
-		/* As we may have blocked we had better restart... */
-		goto restart;
+		/* We hold a reference to 'inode' so it couldn't have been
+		 * removed from s_inodes list while we dropped the inode_lock.
+		 * We cannot iput the inode now as we can be holding the last

fs/dquot.c restart: · line 728

Quoted 10 lines
fs/dquot.c +6 -0
+		 * reference and we cannot iput it under inode_lock. So we
+		 * keep the reference and iput it later. */
+		old_inode = inode;
+		spin_lock(&inode_lock);
 	}
 	spin_unlock(&inode_lock);
+	if (old_inode)
+		iput(old_inode);
 }
 

iput(NULL) is legal. In the great majority of cases old_inode!=NULL so I think we should use that feature here?

fs/dquot.c renamed from fs/dquot.c~quota-improve-inode-list-scanning-in-add_dquot_ref-fix +2 -4
@@ -710,8 +710,7 @@ static void add_dquot_ref(struct super_b
 		__iget(inode);
 		spin_unlock(&inode_lock);
 
-		if (old_inode)
-			iput(old_inode);
+		iput(old_inode);
 		sb->dq_op->initialize(inode, type);
 		/* We hold a reference to 'inode' so it couldn't have been
 		 * removed from s_inodes list while we dropped the inode_lock.
@@ -722,8 +721,7 @@ static void add_dquot_ref(struct super_b
 		spin_lock(&inode_lock);
 	}
 	spin_unlock(&inode_lock);
-	if (old_inode)
-		iput(old_inode);
+	iput(old_inode);
 }
 
 /* Return 0 if dqput() won't block (note that 1 doesn't necessarily mean blocking) */

_

author reviewer
#2 lore raw
Quoted 61 lines · Andrew Morton
Quoted 58 lines · On Tue, 11 Dec 2007 14:43:19 +0100 Jan Kara <[email protected]> wrote:

Hi,

attached patch makes add_dquot_ref() quite faster if used on a life
filesystem. It has survived some testing on my machine but there could be some subtle races I've missed so please leave it in -mm for a while (not that I'd think that people use quotas with -mm kernels but at least it should go into vanilla in -rc1...). Thanks.

Honza

---
We restarted scan of sb->s_inodes list whenever we had to drop inode_lock in add_dquot_ref(). This leads to overall quadratic running time and thus add_dquot_ref() can take several minutes when called on a life filesystem. We fix the problem by using the fact that inode cannot be removed from s_inodes list while we hold a reference to it and thus we can safely restart the scan if we don't drop the reference. Here we use the fact that inodes freshly added to s_inodes list are already guaranteed to have quotas properly initialized and the ordering of inodes on s_inodes list does not change so we cannot skip any inode. Thanks goes to Nick <[email protected]> for analyzing the problem and testing the fix.

fs/dquot.c +10 -5
index 2809768..ecea4be 100644
@@ -696,9 +696,8 @@ static int dqinit_needed(struct inode *inode, int type)
 /* This routine is guarded by dqonoff_mutex mutex */
 static void add_dquot_ref(struct super_block *sb, int type)
 {
-	struct inode *inode;
+	struct inode *inode, *old_inode = NULL;
 
-restart:
 	spin_lock(&inode_lock);
 	list_for_each_entry(inode, &sb->s_inodes, i_sb_list) {
 		if (!atomic_read(&inode->i_writecount))
@@ -711,12 +710,20 @@ restart:
 		__iget(inode);
 		spin_unlock(&inode_lock);
 
+		if (old_inode)
+			iput(old_inode);
 		sb->dq_op->initialize(inode, type);
-		iput(inode);
-		/* As we may have blocked we had better restart... */
-		goto restart;
+		/* We hold a reference to 'inode' so it couldn't have been
+		 * removed from s_inodes list while we dropped the inode_lock.
+		 * We cannot iput the inode now as we can be holding the last
+		 * reference and we cannot iput it under inode_lock. So we
+		 * keep the reference and iput it later. */
+		old_inode = inode;
+		spin_lock(&inode_lock);
 	}
Quoted 8 lines

fs/dquot.c static void add_dquot_ref(struct super_b · line 726

Quoted 5 lines
fs/dquot.c +2 -0
 	spin_unlock(&inode_lock);
+	if (old_inode)
+		iput(old_inode);
 }
 

iput(NULL) is legal. In the great majority of cases old_inode!=NULL so I think we should use that feature here?

OK. Thanks for the improvement.

Quoted 22 lines
fs/dquot.c renamed from fs/dquot.c~quota-improve-inode-list-scanning-in-add_dquot_ref-fix +2 -4
@@ -710,8 +710,7 @@ static void add_dquot_ref(struct super_b
 		__iget(inode);
 		spin_unlock(&inode_lock);
 
-		if (old_inode)
-			iput(old_inode);
+		iput(old_inode);
 		sb->dq_op->initialize(inode, type);
 		/* We hold a reference to 'inode' so it couldn't have been
 		 * removed from s_inodes list while we dropped the inode_lock.
@@ -722,8 +721,7 @@ static void add_dquot_ref(struct super_b
 		spin_lock(&inode_lock);
 	}
 	spin_unlock(&inode_lock);
-	if (old_inode)
-		iput(old_inode);
+	iput(old_inode);
 }
 
 /* Return 0 if dqput() won't block (note that 1 doesn't necessarily mean blocking) */
Quoted 2 lines

_

Honza

Signature · 2 lines
Jan Kara <[email protected]>
SUSE Labs, CR