Re: [Qemu-devel] [PATCH v6 2/2] block/vxhs.c: Add qemu-iotests for new block device type "vxhs"

[Ketan Nilangekar]

On 2/13/17, 8:32 AM, "Jeff Cody" <address@hidden> wrote:

    On Mon, Feb 13, 2017 at 01:37:25PM +0000, Stefan Hajnoczi wrote:
    > On Tue, Feb 07, 2017 at 03:12:36PM -0800, ashish mittal wrote:
    > > On Tue, Nov 8, 2016 at 12:44 PM, Jeff Cody <address@hidden> wrote:
    > > > On Mon, Nov 07, 2016 at 04:59:45PM -0800, Ashish Mittal wrote:
    > > >> These changes use a vxhs test server that is a part of the following
    > > >> repository:
    > > >> https://github.com/MittalAshish/libqnio.git
    > > >>
    > > >> Signed-off-by: Ashish Mittal <address@hidden>
    > > >> ---
    > > >> v6 changelog:
    > > >> (1) Added iotests for VxHS block device.
    > > >>
    > > >>  tests/qemu-iotests/common        |  6 ++++++
    > > >>  tests/qemu-iotests/common.config | 13 +++++++++++++
    > > >>  tests/qemu-iotests/common.filter |  1 +
    > > >>  tests/qemu-iotests/common.rc     | 19 +++++++++++++++++++
    > > >>  4 files changed, 39 insertions(+)
    > > >>
    > > >> diff --git a/tests/qemu-iotests/common b/tests/qemu-iotests/common
    > > >> index d60ea2c..41430d8 100644
    > > >> --- a/tests/qemu-iotests/common
    > > >> +++ b/tests/qemu-iotests/common
    > > >
    > > > When using raw format, I was able to run the test successfully for all
    > > > supported test cases (26 of them).
    > > >
    > > > With qcow2, they fail - but not the fault of this patch, I think; but
    > > > rather, the fault of the test server.  Can qnio_server be modified so 
that
    > > > it does not work on just raw files?
    > > >
    > > >
    > > 
    > > VxHS supports and uses only the raw format.
    > 
    > That's like saying only ext4 guest file systems are supported on VxHS
    > and not ZFS.  The VxHS driver should not care what file system is used,
    > it just does block I/O without interpreting the data.
    > 
    > It must be possible to use any format on top of the VxHS protocol.
    > After all, the image format drivers just do block I/O.  If there is a
    > case where qcow2 on VxHS fails then it needs to be investigated.
    > 
    > The VxHS driver can't be merged until we at least understand the cause
    > of the qcow2 test failures.
    >
    
    A quick run with the test server and a QEMU process showed an abort() in the
    test server, so I just sent a pull req to libqnio to fix that.  
    
    But playing with it in gdb right now with a test qcow2 file, I see that we
    are waiting in aio_poll() forever for the test server to respond to a read
    request, when using qcow2 format.  
    
    As Stefan said, this doesn't really make any sense - why would VXHS behave
    differently based on the file contents?

[Ketan] 
To read/write a qcow2 backed device VxHS server implementation will need to 
understand the qcow2 format. This is not just block IO but actually does 
involve interpreting the qcow2 header and cluster formats. Clearly the test 
server implementation does not handle it as it was never intended to. VxHS 
backend won't handle it either because VxHS virtual disks are written as 
non-sparse files.
There are space savings with the qcow2 format but performance penalties as well 
because of the metadata overhead. As a block storage provider, VxHS does not 
support sparse file formats like qcow2 primarily because of performance reasons.
Implementing a qcow2 backend in the test server would be a non-trivial and 
non-useful exercise since the VxHS server won't support it.

   

 Jeff