On Mon, Aug 19, 2013 at 02:36:14PM +0800, Asias He wrote: > On Fri, Aug 16, 2013 at 10:41:36AM +0200, Stefan Hajnoczi wrote: > > On Mon, Aug 5, 2013 at 10:11 AM, Asias He <as...@redhat.com> wrote: > > A simple but ugly way to fix this is for block.c to also have a > > ->zero_beyond_eof flag which enables the behavior you are adding. > > qcow2_load_vmstate() would disable ->zero_beyond_eof temporarily in > > addition to enabling ->growable. > > I am wondering why the ->growable logic is introduced in the first > place. Adding yet another this kind of flag looks realy ugly ;(
bs->growable serves two functions: 1. It means you can read/write beyond the end of the file, for example, when creating a new image file. Normally BlockDriverState rejects requests beyond the EOF. 2. qcow2 uses it as a hack to implement the vmstate area after the end of regular guest data. This is the ugly part but it always worked up until now. #1 is a legitimate use case. You don't always know how big the file will end up so it's much more convenient to grow on demand, instead of having to bdrv_truncate() all the time. #2 is hacky but hard to solve without duplicating the bounce buffer and coroutine wrapping logic in bdrv_pread() (Kevin has suggested calling bdrv_co_readv() internally instead of bdrv_pread()). Stefan
