Re: [Qemu-devel] [Qemu-block] [PATCH] block/gluster: add support for SEEK_DATA/SEEK_HOLE

[Niels de Vos]

On Tue, Mar 08, 2016 at 01:53:26PM +0100, Kevin Wolf wrote:
> Am 08.03.2016 um 05:21 hat Niels de Vos geschrieben:
> > On Mon, Mar 07, 2016 at 01:27:38PM -0500, Jeff Cody wrote:
> > > On Mon, Mar 07, 2016 at 07:04:15PM +0100, Niels de Vos wrote:
> > > > GlusterFS 3.8 contains support for SEEK_DATA and SEEK_HOLE. This makes
> > > > it possible to detect sparse areas in files.
> > > > 
> > > > Signed-off-by: Niels de Vos <address@hidden>
> > > > 
> > > > --
> > > > Tested by compiling and running "qemu-img map gluster://..." with a
> > > > build of the current master branch of glusterfs. Using a Fedora
> > > > cloud image (in raw format) shows many SEEK procudure calls going back
> > > > and forth over the network. The output of "qemu map" matches the output
> > > > when run against the image on the local filesystem.
> > > > ---
> > > >  block/gluster.c | 159 
> > > > ++++++++++++++++++++++++++++++++++++++++++++++++++++++++
> > > >  configure       |  25 +++++++++
> > > >  2 files changed, 184 insertions(+)
> > > > 
> > > > diff --git a/block/gluster.c b/block/gluster.c
> > > > index 65077a0..1430010 100644
> > > > --- a/block/gluster.c
> > > > +++ b/block/gluster.c
> > > > @@ -677,6 +677,153 @@ static int 
> > > > qemu_gluster_has_zero_init(BlockDriverState *bs)
> > > >      return 0;
> > > >  }
> > > >  
> > > > +#ifdef CONFIG_GLUSTERFS_SEEK_DATA
> > > 
> > > Why do we need to make this a compile-time option?  Version checking
> > > is problematic; for instance, different distributions may have
> > > backported bug fixes / features, that are not reflected by the
> > > reported version number, etc..  Ideally, we can determine
> > > functionality during runtime, and behave accordingly.
> > 
> > This will not get backported to older Gluster versions, it required a
> > protocol change.
> > 
> > > If SEEK_DATA and SEEK_HOLE are not supported,
> > > qemu_gluster_co_get_block_status can return that sectors are all
> > > allocated (which is what happens in block/io.c anyway if the driver
> > > doesn't support the function).
> > 
> > Ok, good to know.
> > 
> > > As long as glfs_lseek() will return error (e.g. EINVAL) for an invalid
> > > whence value,  we can handle it runtime.  Does glfs_lseek() behave
> > > sanely?
> > 
> > Unfortunately older versions of libgfapi do not return EINVAL when
> > SEEK_DATA/HOLE is used. It is something we'll need to fix in the stable
> > releases. We can not assume that all users have installed a version of
> > the library that handles SEEK_DATA/HOLE correctly (return EINVAL) when
> > there is no support in the network protocol or on the server.
> > 
> > To be sure that we don't get some undefined behaviour, the compile time
> > check is needed.
> 
> The code could be compiled on a host with newer libgfapi, but run on a
> different host with an older version. This is why having (only) compile
> time checks is rarely a good idea.

Oh, yes, that is possible. glfs_lseek() is not a new function, so the
symbol version did not need to change.

> Jeff's suggestion to probe the actual behaviour on the host we're
> running on in .bdrv_open() sounds reasonable to me.

Yes, it sure is. I'll send a v2 patch soon.

Thanks,
Niels