Re: [resend PATCH] ext4: avoid deadlock on sync-mounted ext2 FS

[Eric Sandeen <sandeen@redhat.com>]

On 02/10/2012 04:43 PM, Eric Sandeen wrote:
> On 02/10/2012 04:31 AM, Martin Wilck wrote:
>> I don't want to bother, but could someone have a look at this
>> patch please?
> 
> You have to bother, to get attention ;)
> 
>> Martin
>>
>> Processes hang forever on a sync-mounted ext2 file system that
>> is mounted with the ext4 module (default in Fedora 16).
>>
>> I can reproduce this reliably by mounting an ext2 partition with
>> "-o sync" and opening a new file an that partition with vim. vim
>> will hang in "D" state forever.
>>
>> I am attaching a small patch here that solves this issue for me.
>> In the sync mounted case, ext4_handle_dirty_metadata() may call
>> sync_dirty_buffer(), which can't be called with buffer lock held.
>>
>> My knowledge of locking in ext4 is insufficient to judge if this
>> patch is correct, but it may serve as starting point for others.
> 
> This looks an awful lot like:
> 
> b2f49033d80c952a0ffc2d5647bc1a0b8a09c1b3 [PATCH] fix deadlock in ext2
> 
> from 2006!  :)
> 
> But I'm also eyeballing:
> 
> 8a2bfdcbfa441d8b0e5cb9c9a7f45f77f80da465 [PATCH] ext[34]: EA block
> reference count racing fix
> 
> to be sure it doesn't regress anything there.
> 
> And I'd really like to know why this only happens on ext2-with-ext4
> (ext4 nojournal seems ok... hm, actually, I can't repro it on
> ext2 either...)
> 
> Ahah.  Ok, makes sense, it has to be mkfs'd with a 128-byte inode, so
> that selinux xattrs will go into an external block.  Now I can reproduce
> it on ext2.  But still not on ext4, with 128-byte-inodes
> and nojournal.  Odd.  I'd like to sort that out.

I take it back, ext4, 128-byte inodes, -o sync, and ^has_journal does
the trick too.  Ok, good, my world is sane(r) again.


So that means whenever there is no journal, we follow a different path
in __ext4_handle_dirty_metadata(), down to sync_dirty_buffer()

I think moving mb_cache_entry_release() prior to your unlock_buffer
should keep the race closed, like this (not tested):

@@ -487,18 +487,19 @@ ext4_xattr_release_block(handle_t *handle, struct
inode *inode,
                ext4_free_blocks(handle, inode, bh, 0, 1,
                                 EXT4_FREE_BLOCKS_METADATA |
                                 EXT4_FREE_BLOCKS_FORGET);
+               unlock_buffer(bh);
        } else {
                le32_add_cpu(&BHDR(bh)->h_refcount, -1);
+               if (ce)
+                       mb_cache_entry_release(ce);
+               unlock_buffer(bh);
                error = ext4_handle_dirty_metadata(handle, inode, bh);
                if (IS_SYNC(inode))
                        ext4_handle_sync(handle);
                dquot_free_block(inode, 1);
                ea_bdebug(bh, "refcount now=%d; releasing",
                          le32_to_cpu(BHDR(bh)->h_refcount));
-               if (ce)
-                       mb_cache_entry_release(ce);
        }
-       unlock_buffer(bh);
 out:
        ext4_std_error(inode->i_sb, error);
        return;

That's what ext2 does, and what ext3 did, before a bunch of refactoring
a while ago.  Care to send again w/ those 2 lines movedd?

-Eric

> The patch also unlocks prior to mb_cache_entry_release(), and I think
> that might put some of the raciness back?
> 
> -Eric
> 
> 
>> Signed-off-by: Martin.Wilck@ts.fujitsu.com
>> ---
>>  fs/ext4/xattr.c |    3 ++-
>>  1 files changed, 2 insertions(+), 1 deletions(-)
>>
>> diff --git a/fs/ext4/xattr.c b/fs/ext4/xattr.c
>> index 93a00d8..eb771ea 100644
>> --- a/fs/ext4/xattr.c
>> +++ b/fs/ext4/xattr.c
>> @@ -487,8 +487,10 @@ ext4_xattr_release_block(handle_t *handle, struct inode *inode,
>>  		ext4_free_blocks(handle, inode, bh, 0, 1,
>>  				 EXT4_FREE_BLOCKS_METADATA |
>>  				 EXT4_FREE_BLOCKS_FORGET);
>> +		unlock_buffer(bh);
>>  	} else {
>>  		le32_add_cpu(&BHDR(bh)->h_refcount, -1);
>> +		unlock_buffer(bh);
>>  		error = ext4_handle_dirty_metadata(handle, inode, bh);
>>  		if (IS_SYNC(inode))
>>  			ext4_handle_sync(handle);
>> @@ -498,7 +500,6 @@ ext4_xattr_release_block(handle_t *handle, struct inode *inode,
>>  		if (ce)
>>  			mb_cache_entry_release(ce);
>>  	}
>> -	unlock_buffer(bh);
>>  out:
>>  	ext4_std_error(inode->i_sb, error);
>>  	return;
>