Re: Why we need delayed allocation!

On Fri, 2003-12-26 at 23:41, Alex Tomas wrote:
> probably, this would look crazy, but what about following:
> 
> 1) at writeback time we can have in cache dirty pages to be flushed
>    and clean pages we dont flush usually
> 2) if clean page(s) help to enlarge an extent(s) then we could unmap
>    those clean pages and allocate new larger extent using them

The complication with this is O_DIRECT; what if an application does an
O_DIRECT read of the page in parallel to this "remapping"... it could
get the old data from the disk of the new mapping if this happens to be
before the new block is written. Also correct journalling is a tiny bit
complex; eg the metadata blocks should only be updated once the data has
hit the disk in the new spot etc etc (but I'm sure that can be dealt
with).

signature.asc
Description: This is a digitally signed message part