[PATCH] xattr: fine-grained locking

From: Andreas Gruenbacher <agruen@suse.de>

This patch removes the dependency on i_sem in the getxattr and
listxattr iops of ext2 and ext3. In addition, the global ext[23]_xattr
semaphores go away. Instead of i_sem and the global semaphore, mutual
exclusion is now ensured by per-inode xattr semaphores, and by locking
the buffers before modifying them. The detailed locking strategy is
described in comments in fs/ext[23]/xattr.c.

Due to this change it is no longer necessary to take i_sem in
ext[23]_permission() for retrieving acls, so the
ext[23]_permission_locked() functions go away.

Additionally, the patch fixes a race condition in ext[23]_permission:
Accessing inode->i_acl was protected by the BKL in 2.4; in 2.5 there no
longer is such protection. Instead, inode->i_acl (and inode->i_default_acl)
are now accessed under inode->i_lock. (This could be replaced by RCU in
the future.)

In the ext3 extended attribute code, an new uglines results from locking
at the buffer head level: The buffer lock must be held between testing
if an xattr block can be modified and the actual modification to prevent
races from happening. Before a block can be modified,
ext3_journal_get_write_access() must be called. But this requies an unlocked
buffer, so I call ext3_journal_get_write_access() before locking the
buffer. If it turns out that the buffer cannot be modified,
journal_release_buffer() is called. Calling ext3_journal_get_write_access
after the test but while the buffer is still locked would be much better.

1 files changed, 10 insertions, 0 deletions

diff --git a/include/linux/ext3_fs_i.h b/include/linux/ext3_fs_i.h
index 1a6a6c5922f7..1181cfae7142 100644
--- a/include/linux/ext3_fs_i.h
+++ b/include/linux/ext3_fs_i.h
@@ -62,6 +62,16 @@ struct ext3_inode_info {
 	__u32	i_prealloc_count;
 #endif
 	__u32	i_dir_start_lookup;
+#ifdef CONFIG_EXT3_FS_XATTR
+	/*
+	 * Extended attributes can be read independently of the main file
+	 * data. Taking i_sem even when reading would cause contention
+	 * between readers of EAs and writers of regular file data, so
+	 * instead we synchronize on xattr_sem when reading or changing
+	 * EAs.
+	 */
+	struct rw_semaphore xattr_sem;
+#endif
 #ifdef CONFIG_EXT3_FS_POSIX_ACL
 	struct posix_acl	*i_acl;
 	struct posix_acl	*i_default_acl;