btrfs: preserve i_mode if __btrfs_set_acl() fails

When changing a file's acl mask, btrfs_set_acl() will first set the group bits of i_mode to the value of the mask, and only then set the actual extended attribute representing the new acl.

If the second part fails (due to lack of space, for example) and the file had no acl attribute to begin with, the system will from now on assume that the mask permission bits are actual group permission bits, potentially granting access to the wrong users.

Prevent this by starting the journal transaction before calling __btrfs_set_acl and only changing the inode mode after it returns successfully.

---
This issue is covered by generic/449 in xfstests. Several filesystems are affected; some of them have already applied patches:
- fe26569 ext2: preserve i_mode if ext2_set_acl() fails
- f070e5a jfs: preserve i_mode if __jfs_set_acl() fails
- fcea8ae reiserfs: preserve i_mode if __reiserfs_set_acl() fails

 fs/btrfs/acl.c | 29 ++++++++++++++++++++++++++---
 1 file changed, 26 insertions(+), 3 deletions(-)

index 8d8370d..d041526 100644
@@ -27,6 +27,7 @@
 #include "ctree.h"
 #include "btrfs_inode.h"
 #include "xattr.h"
+#include "transaction.h"
 
 struct posix_acl *btrfs_get_acl(struct inode *inode, int type)
 {

@@ -113,14 +114,36 @@ static int __btrfs_set_acl(struct btrfs_trans_handle *trans,
 
 int btrfs_set_acl(struct inode *inode, struct posix_acl *acl, int type)
 {
+	struct btrfs_root *root = BTRFS_I(inode)->root;
+	struct btrfs_trans_handle *trans;
 	int ret;
+	umode_t mode = inode->i_mode;
+
+	if (btrfs_root_readonly(root))
+		return -EROFS;
+
+	trans = btrfs_start_transaction(root, 2);
+	if (IS_ERR(trans))
+		return PTR_ERR(trans);
 
 	if (type == ACL_TYPE_ACCESS && acl) {
-		ret = posix_acl_update_mode(inode, &inode->i_mode, &acl);
+		ret = posix_acl_update_mode(inode, &mode, &acl);
 		if (ret)
-			return ret;
+			goto out;
 	}
-	return __btrfs_set_acl(NULL, inode, acl, type);
+	ret = __btrfs_set_acl(trans, inode, acl, type);
+	if (ret)
+		goto out;
+
+	inode->i_mode = mode;
+	inode_inc_iversion(inode);
+	inode->i_ctime = current_time(inode);
+	set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
+	ret = btrfs_update_inode(trans, root, inode);
+	BUG_ON(ret);
+out:
+	btrfs_end_transaction(trans);
+	return ret;
 }
 
 /*

2.1.4

I had not thought of that, thank you.

If I'm understanding this correctly the issue would be only when setting a NULL default acl on an inode that is not a directory. In that case I probably shouldn't be calling btrfs_update_inode either, but I can't move that back to setxattr.

Perhaps __btrfs_set_acl could return an error in that case, like -ENOTDIR, and then we can set ret back to 0 before returning from btrfs_set_acl.

The call to BUG_ON was already there before my patch, only inside the __btrfs_setxattr function. Since I didn't know the reason I thought it was best not to change it. I'll do as you say in the next version.

Thank you for your review.

When changing a file's acl mask, btrfs_set_acl() will first set the group bits of i_mode to the value of the mask, and only then set the actual extended attribute representing the new acl.

If the second part fails (due to lack of space, for example) and the file had no acl attribute to begin with, the system will from now on assume that the mask permission bits are actual group permission bits, potentially granting access to the wrong users.

Prevent this by starting the journal transaction before calling __btrfs_set_acl and only changing the inode mode after the acl is set successfully.

---
Changes in v2:
- Take the code that checks if we are setting a default acl on something that is not a dir, remove it from the __btrfs_set_acl function, and place it in btrfs_set_acl instead. This should fix the issue pointed out by Josef Bacik, that I was sometimes updating the inode even when there was no change.
- Don't call BUG_ON when the inode failed to update. Also requested by Josef Bacik. It should be noted that __btrfs_setxattr was already calling BUG_ON before my patch; that has not been changed.

 fs/btrfs/acl.c | 31 +++++++++++++++++++++++++++----
 1 file changed, 27 insertions(+), 4 deletions(-)

index 8d8370d..f62e8ac 100644
@@ -27,6 +27,7 @@
 #include "ctree.h"
 #include "btrfs_inode.h"
 #include "xattr.h"
+#include "transaction.h"
 
 struct posix_acl *btrfs_get_acl(struct inode *inode, int type)
 {

@@ -80,8 +81,6 @@ static int __btrfs_set_acl(struct btrfs_trans_handle *trans,
 		name = XATTR_NAME_POSIX_ACL_ACCESS;
 		break;
 	case ACL_TYPE_DEFAULT:
-		if (!S_ISDIR(inode->i_mode))
-			return acl ? -EINVAL : 0;
 		name = XATTR_NAME_POSIX_ACL_DEFAULT;
 		break;
 	default:

@@ -113,14 +112,38 @@ static int __btrfs_set_acl(struct btrfs_trans_handle *trans,
 
 int btrfs_set_acl(struct inode *inode, struct posix_acl *acl, int type)
 {
+	struct btrfs_root *root = BTRFS_I(inode)->root;
+	struct btrfs_trans_handle *trans;
 	int ret;
+	umode_t mode = inode->i_mode;
 
+	if (type == ACL_TYPE_DEFAULT && !S_ISDIR(inode->i_mode))
+		return acl ? -EINVAL : 0;
 	if (type == ACL_TYPE_ACCESS && acl) {
-		ret = posix_acl_update_mode(inode, &inode->i_mode, &acl);
+		ret = posix_acl_update_mode(inode, &mode, &acl);
 		if (ret)
 			return ret;
 	}
-	return __btrfs_set_acl(NULL, inode, acl, type);
+
+	if (btrfs_root_readonly(root))
+		return -EROFS;
+
+	trans = btrfs_start_transaction(root, 2);
+	if (IS_ERR(trans))
+		return PTR_ERR(trans);
+
+	ret = __btrfs_set_acl(trans, inode, acl, type);
+	if (ret)
+		goto out;
+
+	inode->i_mode = mode;
+	inode_inc_iversion(inode);
+	inode->i_ctime = current_time(inode);
+	set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
+	ret = btrfs_update_inode(trans, root, inode);
+out:
+	btrfs_end_transaction(trans);
+	return ret;
 }
 
 /*

2.1.4

When changing a file's acl mask, btrfs_set_acl() will first set the group bits of i_mode to the value of the mask, and only then set the actual extended attribute representing the new acl.

If the second part fails (due to lack of space, for example) and the file had no acl attribute to begin with, the system will from now on assume that the mask permission bits are actual group permission bits, potentially granting access to the wrong users.

Prevent this by restoring the original mode bits if __btrfs_set_acl fails.

---
Please ignore the two previous versions, this is far simpler and has the same effect. To Josef Bacik: thank you for your review, I'm sorry I wasted your time.

 fs/btrfs/acl.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

index 8d8370d..1ba49eb 100644
@@ -114,13 +114,17 @@ static int __btrfs_set_acl(struct btrfs_trans_handle *trans,
 int btrfs_set_acl(struct inode *inode, struct posix_acl *acl, int type)
 {
 	int ret;
+	umode_t old_mode = inode->i_mode;
 
 	if (type == ACL_TYPE_ACCESS && acl) {
 		ret = posix_acl_update_mode(inode, &inode->i_mode, &acl);
 		if (ret)
 			return ret;
 	}
-	return __btrfs_set_acl(NULL, inode, acl, type);
+	ret = __btrfs_set_acl(NULL, inode, acl, type);
+	if (ret)
+		inode->i_mode = old_mode;
+	return ret;
 }
 
 /*

2.1.4

This version is much better, starting the transaction would add some overhead and would need to be measured so we have an idea about the impact.