Skip to contentSkip to conversation

btrfs: move ulist allocation out of transaction in quota enable

PATCH Superseded David Sterba 3 participants 5 messages Feb 25, 2019 – Feb 27, 2019 Btrfs file system

lore
Share & tools
lore

David Sterba

author maintainer
#0 lore raw

PATCH 1/1 btrfs: move ulist allocation out of transaction in quota enable

The allocation happens with GFP_KERNEL after a transaction has been started, this can potentially cause deadlock if reclaim tries to get the memory by flushing filesystem data.

The qgroup_ulist is not used during transaction start as the quotas are not yet enabled but to provide the same semantics, use a temporary variable and assign fs_info::qgroup_ulist as before.

---
 fs/btrfs/qgroup.c | 14 ++++++++------
 1 file changed, 8 insertions(+), 6 deletions(-)
fs/btrfs/qgroup.c +8 -6
index c1cd5558a646..9e49d5e132b8 100644
@@ -887,6 +887,7 @@ int btrfs_quota_enable(struct btrfs_fs_info *fs_info)
 	struct btrfs_key found_key;
 	struct btrfs_qgroup *qgroup = NULL;
 	struct btrfs_trans_handle *trans = NULL;
+	struct ulist *tmp_ulist;
 	int ret = 0;
 	int slot;
 
@@ -894,6 +895,12 @@ int btrfs_quota_enable(struct btrfs_fs_info *fs_info)
 	if (fs_info->quota_root)
 		goto out;
 
+	tmp_ulist = ulist_alloc(GFP_KERNEL);
+	if (!tmp_ulist) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
 	/*
 	 * 1 for quota root item
 	 * 1 for BTRFS_QGROUP_STATUS item
@@ -909,12 +916,7 @@ int btrfs_quota_enable(struct btrfs_fs_info *fs_info)
 		goto out;
 	}
 
-	fs_info->qgroup_ulist = ulist_alloc(GFP_KERNEL);
-	if (!fs_info->qgroup_ulist) {
-		ret = -ENOMEM;
-		btrfs_abort_transaction(trans, ret);
-		goto out;
-	}
+	fs_info->qgroup_ulist = tmp_ulist;
 
 	/*
 	 * initially create the quota tree
Signature · 1 line
2.20.1
#1 lore raw
Quoted 27 lines · David Sterba

The allocation happens with GFP_KERNEL after a transaction has been started, this can potentially cause deadlock if reclaim tries to get the memory by flushing filesystem data.

The qgroup_ulist is not used during transaction start as the quotas are not yet enabled but to provide the same semantics, use a temporary variable and assign fs_info::qgroup_ulist as before.

---
 fs/btrfs/qgroup.c | 14 ++++++++------
 1 file changed, 8 insertions(+), 6 deletions(-)
fs/btrfs/qgroup.c +1 -0
index c1cd5558a646..9e49d5e132b8 100644
@@ -887,6 +887,7 @@ int btrfs_quota_enable(struct btrfs_fs_info *fs_info)
        struct btrfs_key found_key;
        struct btrfs_qgroup *qgroup = NULL;
        struct btrfs_trans_handle *trans = NULL;
+       struct ulist *tmp_ulist;
        int ret = 0;
        int slot;
 

fs/btrfs/qgroup.c int btrfs_quota_enable(struct btrfs_fs_info *fs_info) · line 902

Quoted 8 lines
fs/btrfs/qgroup.c +5 -0
@@ -894,6 +895,12 @@ int btrfs_quota_enable(struct btrfs_fs_info *fs_info)
        if (fs_info->quota_root)
                goto out;
 
+       tmp_ulist = ulist_alloc(GFP_KERNEL);
+       if (!tmp_ulist) {
+               ret = -ENOMEM;
+               goto out;
+       }

tmp_ulist needs to be freed if starting the transaction below fails.

Other then that, looks good.

Quoted 21 lines
fs/btrfs/qgroup.c +2 -7
+
        /*
         * 1 for quota root item
         * 1 for BTRFS_QGROUP_STATUS item
@@ -909,12 +916,7 @@ int btrfs_quota_enable(struct btrfs_fs_info *fs_info)
                goto out;
        }
 
-       fs_info->qgroup_ulist = ulist_alloc(GFP_KERNEL);
-       if (!fs_info->qgroup_ulist) {
-               ret = -ENOMEM;
-               btrfs_abort_transaction(trans, ret);
-               goto out;
-       }
+       fs_info->qgroup_ulist = tmp_ulist;
 
        /*
         * initially create the quota tree
--

2.20.1

Signature · 3 lines
Filipe David Manana,

“Whether you think you can, or you think you can't — you're right.”
#2 lore raw
Quoted 31 lines · Filipe Manana
Quoted 30 lines · David Sterba

The allocation happens with GFP_KERNEL after a transaction has been started, this can potentially cause deadlock if reclaim tries to get the memory by flushing filesystem data.

The qgroup_ulist is not used during transaction start as the quotas are not yet enabled but to provide the same semantics, use a temporary variable and assign fs_info::qgroup_ulist as before.

---
 fs/btrfs/qgroup.c | 14 ++++++++------
 1 file changed, 8 insertions(+), 6 deletions(-)
fs/btrfs/qgroup.c +1 -0
index c1cd5558a646..9e49d5e132b8 100644
@@ -887,6 +887,7 @@ int btrfs_quota_enable(struct btrfs_fs_info *fs_info)
        struct btrfs_key found_key;
        struct btrfs_qgroup *qgroup = NULL;
        struct btrfs_trans_handle *trans = NULL;
+       struct ulist *tmp_ulist;
        int ret = 0;
        int slot;
 
@@ -894,6 +895,12 @@ int btrfs_quota_enable(struct btrfs_fs_info *fs_info)
        if (fs_info->quota_root)
                goto out;
 
Quoted 7 lines

fs/btrfs/qgroup.c int btrfs_quota_enable(struct btrfs_fs_info *fs_info) · line 902

Quoted 5 lines
fs/btrfs/qgroup.c +5 -0
+       tmp_ulist = ulist_alloc(GFP_KERNEL);
+       if (!tmp_ulist) {
+               ret = -ENOMEM;
+               goto out;
+       }

tmp_ulist needs to be freed if starting the transaction below fails.

Right. If I did not try to preserve the semantics (that does not seem to be necessary), the freeing would happen on qgroup_ulist as before.

Starting transaction uses qgruop_ulist if quotas are enabled, so at this time of btrfs_quota_enable, this is not yet true. The status bit is set later.

I'll update the patch with this explanation and drop the temporary variable.

author maintainer
#3 lore raw

PATCH v2 1/1 btrfs: move ulist allocation out of transaction in quota enable

In reply to David Sterba

The allocation happens with GFP_KERNEL after a transaction has been started, this can potentially cause deadlock if reclaim tries to get the memory by flushing filesystem data.

The fs_info::qgroup_ulist is not used during transaction start when quotas are not enabled. The status bit BTRFS_FS_QUOTA_ENABLED is set later in btrfs_quota_enable so it's safe to move it before the transaction start.

v2:
- use qgroup_ulist directly, drop the temporary variable
- update changelog

---
 fs/btrfs/qgroup.c | 13 ++++++-------
 1 file changed, 6 insertions(+), 7 deletions(-)
fs/btrfs/qgroup.c +6 -7
index c1cd5558a646..eb680b715dd6 100644
@@ -894,6 +894,12 @@ int btrfs_quota_enable(struct btrfs_fs_info *fs_info)
 	if (fs_info->quota_root)
 		goto out;
 
+	fs_info->qgroup_ulist = ulist_alloc(GFP_KERNEL);
+	if (!fs_info->qgroup_ulist) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
 	/*
 	 * 1 for quota root item
 	 * 1 for BTRFS_QGROUP_STATUS item
@@ -909,13 +915,6 @@ int btrfs_quota_enable(struct btrfs_fs_info *fs_info)
 		goto out;
 	}
 
-	fs_info->qgroup_ulist = ulist_alloc(GFP_KERNEL);
-	if (!fs_info->qgroup_ulist) {
-		ret = -ENOMEM;
-		btrfs_abort_transaction(trans, ret);
-		goto out;
-	}
-
 	/*
 	 * initially create the quota tree
 	 */
Signature · 1 line
2.20.1
Reviewed-by on v2 1/1
#4 lore raw
Quoted 11 lines · David Sterba

The allocation happens with GFP_KERNEL after a transaction has been started, this can potentially cause deadlock if reclaim tries to get the memory by flushing filesystem data.

The fs_info::qgroup_ulist is not used during transaction start when quotas are not enabled. The status bit BTRFS_FS_QUOTA_ENABLED is set later in btrfs_quota_enable so it's safe to move it before the transaction start.

Looks good now.

Quoted 42 lines

v2:
- use qgroup_ulist directly, drop the temporary variable
- update changelog

---
 fs/btrfs/qgroup.c | 13 ++++++-------
 1 file changed, 6 insertions(+), 7 deletions(-)
fs/btrfs/qgroup.c +6 -8
index c1cd5558a646..eb680b715dd6 100644
@@ -894,6 +894,12 @@ int btrfs_quota_enable(struct btrfs_fs_info *fs_info)
        if (fs_info->quota_root)
                goto out;
 
+       fs_info->qgroup_ulist = ulist_alloc(GFP_KERNEL);
+       if (!fs_info->qgroup_ulist) {
+               ret = -ENOMEM;
+               goto out;
+       }
+
        /*
         * 1 for quota root item
         * 1 for BTRFS_QGROUP_STATUS item
@@ -909,13 +915,6 @@ int btrfs_quota_enable(struct btrfs_fs_info *fs_info)
                goto out;
        }
 
-       fs_info->qgroup_ulist = ulist_alloc(GFP_KERNEL);
-       if (!fs_info->qgroup_ulist) {
-               ret = -ENOMEM;
-               btrfs_abort_transaction(trans, ret);
-               goto out;
-       }
-
        /*
         * initially create the quota tree
         */
--

2.20.1

Signature · 3 lines
Filipe David Manana,

“Whether you think you can, or you think you can't — you're right.”