Re: [PATCH] ext4: ext4_split_extent shoult take care about extent zeroout

[Dmitry Monakhov <dmonakhov@openvz.org>]

On Thu, 31 Jan 2013 15:18:34 +0100, Jan Kara <jack@suse.cz> wrote:
> On Thu 31-01-13 11:24:58, Dmitry Monakhov wrote:
> > We have to update extent's state after first ext4_split_extent_at otherwise this result
> > in following trace:
> > ->ext4_ext_handle_uninitialized_extents (ex=[1000:20:uninit], lblock 1000, max_blocks 10)
> >   ->ext4_split_extent_at(ex=[1000,128], lblk 10010) /// First split
> >     ->ext4_ext_split() -> ENOSPC
> >     ->ext4_ext_zeroout
> >       ->ext4_ext_dirty  -> ex=[1000:20:init]
> >   ->ext4_split_extent_at(ex=[1000,128], lblk 10000) /// Second split
> >      if(split == ee_block)
> >          if (split_flag & EXT4_EXT_MARK_UNINIT2)
> >             ext4_ext_mark_uninitialized(ex); ex=[1000:20:uninit] /// The bug!
> >      ->ext4_ext_dirty ->ex=[1000:20:uninit]
> > 
> > At the end ext4_convert_unwritten_extents_endio() will findout large uninitialized
> > extent.
>   Thanks for debugging this. You fix look correct so you can add
> Reviewed-by: Jan Kara <jack@suse.cz>
Actually patch itself is sub-optimal because second split probably will
also hit ENOSPC and endup with second zeroout. Off course ENOSPC is not
the place where optimization should take place but still.
I've already send updated version here:
message-id:<1359643738-22435-1-git-send-email-dmonakhov@openvz.org>

>   but I have to say the above changelog isn't optimal. I had to look into
> the code to verify you are actually speaking about what I think you are
> speaking. I think there are some mistakes in block numbers and notation for
> extents isn't comletely clear either. Can we make the changelog something
> like:
> 
> When ext4_split_extent_at() ends up doing zeroout & conversion to
> initialized instead of split & conversion, ext4_split_extent() gets
> confused and can wrongly mark the extent back as uninitialized resulting in
> end IO code getting confused from large unwritten extents (it doesn't
> result in data corruption mostly by luck).
But it is likely to result in data loss because xxx_endio also probably
failed to split extent due to ENOSPC.
> 
> The example of problematic behavior is:
>                            lblk len              lblk len
>   ext4_split_extent() (ex=[1000,30,uninit], map=[1010,10])
>     ext4_split_extent_at() (split [1000,30,uninit] at 1020)
>       ext4_ext_insert_extent() -> ENOSPC
>       ext4_ext_zeroout()
>         -> extent [1000,30] is now initialized
>     ext4_split_extent_at() (split [1000,30,init] at 1010,
> 			    MARK_UNINIT1 | MARK_UNINIT2)
>       -> extent is split and parts marked as uninitialized
> 
> Fix the problem by rechecking extent type after the first
> ext4_split_extent_at() returns.
> ---
> 
> What do you think? BTW: we don't have to further try to split the extent
> once it gets initialized do we? For now I'd keep your fix just to make
> ext4_split_extent() generic but noone really calls that function for
> initialized extent or is interested in splitting once the extent gets
> initialized. That code seriously needs some diet...
That is correct, but it is not illegal to split initialized extents
(punch_hole theoretically may use that). But what I'm absolutely sure that
(MARK_UNINIT1|MARK_UNINIT2|MAY_ZEROOUT) flags is not applicable to
initialized extent (see new bugon in second version of my patch)
> 
> 								Honza
> 
> > 
> > TESTCASE: https://github.com/dmonakhov/xfstests/commit/1a1c4f337d4d198803436c63a56625b1a78d8a5e
> > 
> > Signed-off-by: Dmitry Monakhov <dmonakhov@openvz.org>
> > ---
> >  fs/ext4/extents.c |   14 ++++++++++----
> >  1 files changed, 10 insertions(+), 4 deletions(-)
> > 
> > diff --git a/fs/ext4/extents.c b/fs/ext4/extents.c
> > index 97cac01..7a3f679 100644
> > --- a/fs/ext4/extents.c
> > +++ b/fs/ext4/extents.c
> > @@ -3091,18 +3091,24 @@ static int ext4_split_extent(handle_t *handle,
> >  		if (err)
> >  			goto out;
> >  	}
> > -
> > +	/* Update path is required because previous ext4_split_extent_at() may
> > +	 * result in split of original leaf or extent zeroout.
> > +	 */
>   Style nit comment should look like:
> 	/*
> 	 * Update path is required because previous ext4_split_extent_at() may
> 	 * result in split of original leaf or extent zeroout.
> 	 */
> 
> >  	ext4_ext_drop_refs(path);
> >  	path = ext4_ext_find_extent(inode, map->m_lblk, path);
> >  	if (IS_ERR(path))
> >  		return PTR_ERR(path);
> > +	depth = ext_depth(inode);
> > +	ex = path[depth].p_ext;
> > +	uninitialized = ext4_ext_is_uninitialized(ex);
> >  
> >  	if (map->m_lblk >= ee_block) {
> >  		split_flag1 = split_flag & EXT4_EXT_MAY_ZEROOUT;
> > -		if (uninitialized)
> > +		if (uninitialized) {
> >  			split_flag1 |= EXT4_EXT_MARK_UNINIT1;
> > -		if (split_flag & EXT4_EXT_MARK_UNINIT2)
> > -			split_flag1 |= EXT4_EXT_MARK_UNINIT2;
> > +			if (split_flag & EXT4_EXT_MARK_UNINIT2)
> > +				split_flag1 |= EXT4_EXT_MARK_UNINIT2;
> > +		}
> >  		err = ext4_split_extent_at(handle, inode, path,
> >  				map->m_lblk, split_flag1, flags);
> >  		if (err)
> > -- 
> > 1.7.1
> > 
> -- 
> Jan Kara <jack@suse.cz>
> SUSE Labs, CR
> --
> To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html