On Mon, Jul 11, 2016 at 03:06:03PM +0800, Kaho Ng wrote:
> Just wonders why we prefer failing the request of refilling freelist
> with XFS_WANT_CORRUPTED_RETURN(mp, i == 1) in some rare case, rather
> than returning NULLAGBLOCK and allowing the loop in
> xfs_alloc_ag_vextent_size() to try xfs_alloc_ag_vextent_small()...
Have a look at where xfs_alloc_ag_vextent_small() gets the blocks it
returns to the caller if the btree cursor doesn't point to a btree
record we can use. i.e. you can't refill the free list from
xfs_alloc_ag_vextent_small() because it allocates blocks from ...
> In
> such corner case there will always be a lot of small extents at the
> front of the by-count tree, and any truncation changes to the first
> entry in the tree will not result in tree splits and triggering
> assertion failure.
If there are records we can use, then we'll allocate them from the
btree. Failure to allocate from the btree indicates something is
inconsistent, there's a bug in the code or we've got corruption
occuring. A corruption shutdown is the only safe course of action
when we find something confusing like this - if we guess wrong them
we'll only make the bad state/corruption worse.
Cheers,
Dave.
--
Dave Chinner
david@xxxxxxxxxxxxx
|