ext4: avoid lockdep warning when inheriting encryption context
authorEric Biggers <ebiggers@google.com>
Mon, 21 Nov 2016 16:52:44 +0000 (11:52 -0500)
committerTheodore Ts'o <tytso@mit.edu>
Mon, 21 Nov 2016 16:52:44 +0000 (11:52 -0500)
On a lockdep-enabled kernel, xfstests generic/027 fails due to a lockdep
warning when run on ext4 mounted with -o test_dummy_encryption:

    xfs_io/4594 is trying to acquire lock:
     (jbd2_handle
    ){++++.+}, at:
    [<ffffffff813096ef>] jbd2_log_wait_commit+0x5/0x11b

    but task is already holding lock:
     (jbd2_handle
    ){++++.+}, at:
    [<ffffffff813000de>] start_this_handle+0x354/0x3d8

The abbreviated call stack is:

 [<ffffffff813096ef>] ? jbd2_log_wait_commit+0x5/0x11b
 [<ffffffff8130972a>] jbd2_log_wait_commit+0x40/0x11b
 [<ffffffff813096ef>] ? jbd2_log_wait_commit+0x5/0x11b
 [<ffffffff8130987b>] ? __jbd2_journal_force_commit+0x76/0xa6
 [<ffffffff81309896>] __jbd2_journal_force_commit+0x91/0xa6
 [<ffffffff813098b9>] jbd2_journal_force_commit_nested+0xe/0x18
 [<ffffffff812a6049>] ext4_should_retry_alloc+0x72/0x79
 [<ffffffff812f0c1f>] ext4_xattr_set+0xef/0x11f
 [<ffffffff812cc35b>] ext4_set_context+0x3a/0x16b
 [<ffffffff81258123>] fscrypt_inherit_context+0xe3/0x103
 [<ffffffff812ab611>] __ext4_new_inode+0x12dc/0x153a
 [<ffffffff812bd371>] ext4_create+0xb7/0x161

When a file is created in an encrypted directory, ext4_set_context() is
called to set an encryption context on the new file.  This calls
ext4_xattr_set(), which contains a retry loop where the journal is
forced to commit if an ENOSPC error is encountered.

If the task actually were to wait for the journal to commit in this
case, then it would deadlock because a handle remains open from
__ext4_new_inode(), so the running transaction can't be committed yet.
Fortunately, __jbd2_journal_force_commit() avoids the deadlock by not
allowing the running transaction to be committed while the current task
has it open.  However, the above lockdep warning is still triggered.

This was a false positive which was introduced by: 1eaa566d368b: jbd2:
track more dependencies on transaction commit

Fix the problem by passing the handle through the 'fs_data' argument to
ext4_set_context(), then using ext4_xattr_set_handle() instead of
ext4_xattr_set().  And in the case where no journal handle is specified
and ext4_set_context() has to open one, add an ENOSPC retry loop since
in that case it is the outermost transaction.

Signed-off-by: Eric Biggers <ebiggers@google.com>
fs/ext4/ialloc.c
fs/ext4/super.c

index 088afe07ddda24b7fe1f2a42e9674679941d1f73..e57e8d90ea54a71745cf6257a7f1ef6437a0191b 100644 (file)
@@ -1115,8 +1115,7 @@ got:
        }
 
        if (encrypt) {
-               /* give pointer to avoid set_context with journal ops. */
-               err = fscrypt_inherit_context(dir, inode, &encrypt, true);
+               err = fscrypt_inherit_context(dir, inode, handle, true);
                if (err)
                        goto fail_free_drop;
        }
index cbfaee175fa5eeb43872f292c90cefe9626023ea..bb3a8edc75db94c2d4c6079f045f75ede54cd69f 100644 (file)
@@ -1114,14 +1114,22 @@ static int ext4_prepare_context(struct inode *inode)
 static int ext4_set_context(struct inode *inode, const void *ctx, size_t len,
                                                        void *fs_data)
 {
-       handle_t *handle;
-       int res, res2;
+       handle_t *handle = fs_data;
+       int res, res2, retries = 0;
 
-       /* fs_data is null when internally used. */
-       if (fs_data) {
-               res  = ext4_xattr_set(inode, EXT4_XATTR_INDEX_ENCRYPTION,
-                               EXT4_XATTR_NAME_ENCRYPTION_CONTEXT, ctx,
-                               len, 0);
+       /*
+        * If a journal handle was specified, then the encryption context is
+        * being set on a new inode via inheritance and is part of a larger
+        * transaction to create the inode.  Otherwise the encryption context is
+        * being set on an existing inode in its own transaction.  Only in the
+        * latter case should the "retry on ENOSPC" logic be used.
+        */
+
+       if (handle) {
+               res = ext4_xattr_set_handle(handle, inode,
+                                           EXT4_XATTR_INDEX_ENCRYPTION,
+                                           EXT4_XATTR_NAME_ENCRYPTION_CONTEXT,
+                                           ctx, len, 0);
                if (!res) {
                        ext4_set_inode_flag(inode, EXT4_INODE_ENCRYPT);
                        ext4_clear_inode_state(inode,
@@ -1134,14 +1142,15 @@ static int ext4_set_context(struct inode *inode, const void *ctx, size_t len,
                return res;
        }
 
+retry:
        handle = ext4_journal_start(inode, EXT4_HT_MISC,
                        ext4_jbd2_credits_xattr(inode));
        if (IS_ERR(handle))
                return PTR_ERR(handle);
 
-       res = ext4_xattr_set(inode, EXT4_XATTR_INDEX_ENCRYPTION,
-                       EXT4_XATTR_NAME_ENCRYPTION_CONTEXT, ctx,
-                       len, 0);
+       res = ext4_xattr_set_handle(handle, inode, EXT4_XATTR_INDEX_ENCRYPTION,
+                                   EXT4_XATTR_NAME_ENCRYPTION_CONTEXT,
+                                   ctx, len, 0);
        if (!res) {
                ext4_set_inode_flag(inode, EXT4_INODE_ENCRYPT);
                /* Update inode->i_flags - e.g. S_DAX may get disabled */
@@ -1151,6 +1160,9 @@ static int ext4_set_context(struct inode *inode, const void *ctx, size_t len,
                        EXT4_ERROR_INODE(inode, "Failed to mark inode dirty");
        }
        res2 = ext4_journal_stop(handle);
+
+       if (res == -ENOSPC && ext4_should_retry_alloc(inode->i_sb, &retries))
+               goto retry;
        if (!res)
                res = res2;
        return res;