Re: [PATCH 8/7] xfs: delalloc -> unwritten COW fork allocation can go wrong

[Brian Foster <bfoster@redhat.com>]

On Tue, Nov 20, 2018 at 05:36:05PM +1100, Dave Chinner wrote:
> 
> From: Dave Chinner <dchinner@redhat.com>
> 
> Long saga. There have been days spent following this through dead end
> after dead end in multi-GB event traces. This morning, after writing
> a trace-cmd wrapper that enabled me to be more selective about XFS
> trace points, I discovered that I could get just enough essential
> tracepoints enabled that there was a 50:50 chance the fsx config
> would fail at ~115k ops. If it didn't fail at op 115547, I stopped
> fsx at op 115548 anyway.
> 
> That gave me two traces - one where the problem manifested, and one
> where it didn't. After refining the traces to have the necessary
> information, I found that in the failing case there was a real
> extent in the COW fork compared to an unwritten extent in the
> working case.
> 
> Walking back through the two traces to the point where the CWO fork
> extents actually diverged, I found that the bad case had an extra
> unwritten extent in it. This is likely because the bug it led me to
> had triggered multiple times in those 115k ops, leaving stray
> COW extents around. What I saw was a COW delalloc conversion to an
> unwritten extent (as they should always be through
> xfs_iomap_write_allocate()) resulted in a /written extent/:
> 
> xfs_writepage:        dev 259:0 ino 0x83 pgoff 0x17000 size 0x79a00 offset 0 length 0
> xfs_iext_remove:      dev 259:0 ino 0x83 state RC|LF|RF|COW cur 0xffff888247b899c0/2 offset 32 block 152 count 20 flag 1 caller xfs_bmap_add_extent_delay_real
> xfs_bmap_pre_update:  dev 259:0 ino 0x83 state RC|LF|RF|COW cur 0xffff888247b899c0/1 offset 1 block 4503599627239429 count 31 flag 0 caller xfs_bmap_add_extent_delay_real
> xfs_bmap_post_update: dev 259:0 ino 0x83 state RC|LF|RF|COW cur 0xffff888247b899c0/1 offset 1 block 121 count 51 flag 0 caller xfs_bmap_add_ex
> 
> Basically, Cow fork before:
> 
> 	0 1            32          52
> 	+H+DDDDDDDDDDDD+UUUUUUUUUUU+
> 	   PREV		RIGHT
> 
> COW delalloc conversion allocates:
> 
> 	  1	       32
> 	  +uuuuuuuuuuuu+
> 	  NEW
> 
> And the result according to the xfs_bmap_post_update trace was:
> 
> 	0 1            32          52
> 	+H+wwwwwwwwwwwwwwwwwwwwwwww+
> 	   PREV
> 
> Which is clearly wrong - it should be a merged unwritten extent,
> not an unwritten extent.
> 
> That lead me to look at the LEFT_FILLING|RIGHT_FILLING|RIGHT_CONTIG
> case in xfs_bmap_add_extent_delay_real(), and sure enough, there's
> the bug.
> 
> It takes the old delalloc extent (PREV) and adds the length of the
> RIGHT extent to it, takes the start block from NEW, removes the
> RIGHT extent and then updates PREV with the new extent.
> 
> What it fails to do is update PREV.br_state. For delalloc, this is
> always XFS_EXT_NORM, while in this case we are converting the
> delayed allocation to unwritten, so it needs to be updated to
> XFS_EXT_UNWRITTEN. This LF|RF|RC case does not do this, and so
> the resultant extent is always written.
> 
> And that's the bug I've been chasing for a week - a bmap btree bug,
> not a reflink/dedupe/copy_file_range bug, but a BMBT bug introduced
> with the recent in core extent tree scalability enhancements.
> 
> Signed-off-by: Dave Chinner <dchinner@redhat.com>
> ---
>  fs/xfs/libxfs/xfs_bmap.c | 5 ++++-
>  1 file changed, 4 insertions(+), 1 deletion(-)
> 
> diff --git a/fs/xfs/libxfs/xfs_bmap.c b/fs/xfs/libxfs/xfs_bmap.c
> index 2d78fd53e822..39eaa2b86060 100644
> --- a/fs/xfs/libxfs/xfs_bmap.c
> +++ b/fs/xfs/libxfs/xfs_bmap.c
> @@ -1694,10 +1694,13 @@ xfs_bmap_add_extent_delay_real(
>  	case BMAP_LEFT_FILLING | BMAP_RIGHT_FILLING | BMAP_RIGHT_CONTIG:
>  		/*
>  		 * Filling in all of a previously delayed allocation extent.
> -		 * The right neighbor is contiguous, the left is not.
> +		 * The right neighbor is contiguous, the left is not. Take care
> +		 * with delay -> unwritten extent allocation here because the
> +		 * delalloc record we are overwriting is always written.
>  		 */
>  		PREV.br_startblock = new->br_startblock;
>  		PREV.br_blockcount += RIGHT.br_blockcount;
> +		PREV.br_state = new->br_state;
>  
>  		xfs_iext_next(ifp, &bma->icur);
>  		xfs_iext_remove(bma->ip, &bma->icur, state);

Fix looks sane to me, though I'm a little curious why this doesn't
follow the "contiguous extent extension" pattern that most of the other
contig cases seem to follow. For example, something like the following
for the right contig case:

                old = RIGHT;
                RIGHT.br_startoff = new->br_startoff;
                RIGHT.br_startblock = new->br_startblock;
                RIGHT.br_blockcount += new.br_blockcount;

                xfs_iext_remove(bma->ip, &bma->icur, state); /* PREV */
                xfs_iext_next(ifp, &bma->icur);
                xfs_iext_update_extent(bma->ip, state, &bma->icur, &RIGHT);

... and change the subsequent btree update to use old/RIGHT. Maybe
Christoph has thoughts on that.

Brian