Re: [Qemu-devel] blkdebug and VMDK (iotests 033 failure on monolithicFlat)

[Fam Zheng <famz@redhat.com>]

On Wed, 05/06 16:12, Max Reitz wrote:
> On 06.05.2015 15:59, Fam Zheng wrote:
> >On Wed, 05/06 15:45, Max Reitz wrote:
> >>On 05.05.2015 12:44, Fam Zheng wrote:
> >>>Max,
> >>>
> >>>Since you once fixed VMDK with the json descriptor filename, could you take a
> >>>look at the error:
> >>>
> >>>$ ./check -vmdk -o "subformat=monolithicFlat" 033
> >>>
> >>>033         - output mismatch (see 033.out.bad)
> >>>--- /home/fam/qemu/tests/qemu-iotests/033.out   2015-05-05 10:52:50.524378312 +0800
> >>>+++ 033.out.bad 2015-05-05 17:38:36.147369924 +0800
> >>>@@ -2,54 +2,36 @@
> >>>  Formatting 'TEST_DIR/t.IMGFMT', fmt=IMGFMT size=134217728
> >>>  == preparing image ==
> >>>-wrote 1024/1024 bytes at offset 512
> >>>-1 KiB, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec)
> >>>-wrote 1536/1536 bytes at offset 131072
> >>>-1.500 KiB, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec)
> >>>-wrote 131072/131072 bytes at offset 1024
> >>>-128 KiB, X ops; XX:XX:XX.X (XXX YYY/sec and XXX ops/sec)
> >>>+qemu-io: can't open device blkdebug::/tmp/qemu-iotests/t.vmdk: Cannot use relative extent paths with VMDK descriptor file 'json:{"image": {"driver": "file", "filename": "/tmp/qemu-iotests/t.vmdk"}, "driver": "blkdebug", "align": "512"}'
> >>>+qemu-io: can't open device blkdebug::/tmp/qemu-iotests/t.vmdk: Cannot use relative extent paths with VMDK descriptor file 'json:{"image": {"driver": "file", "filename": "/tmp/qemu-iotests/t.vmdk"}, "driver": "blkdebug", "align": "512"}'
> >>>+qemu-io: can't open device blkdebug::/tmp/qemu-iotests/t.vmdk: Cannot use relative extent paths with VMDK descriptor file 'json:{"image": {"driver": "file", "filename": "/tmp/qemu-iotests/t.vmdk"}, "driver": "blkdebug", "align": "512"}'
> >>>
> >>>...
> >>>
> >>>
> >>>I'm not sure what's the best fix here. It's possible to dig out
> >>>bs->file->file->filename in this case, but I'm not sure what's the rule of this
> >>>filename mystery.
> >>>
> >>>Fam
> >>Hi Fam,
> >>
> >>The problem is that the align option is (consciously) set for blkdebug, so
> >>the blkdebug driver cannot just convert it to a blkdebug::$foo filename, but
> >>must create a JSON filename so that the align option is included.
> >>
> >>I seem to recall that someone (me? I don't know) once proposed a way for
> >>block drivers not only to present a file name (which is json:{${baz}} in
> >>this case), but also the name of the directory they are in, or better, a
> >>template to which a relative path can be appended (in Unix syntax, that is,
> >>with / as separators and .. and so on). The difference to the file name
> >>would be that we can ignore all the options given to the block driver for
> >>controlling its behavior, we just need a path name.
> >>
> >>I don't know why we didn't pursue that proposal, but it does seem to be
> >>necessary here.
> >>
> >>I think we can integrate it into the existing bdrv_refresh_filename()
> >>"infrastructure", so it shouldn't be too difficult. I can take a shot at it,
> >>if you agree.
> >>
> >Sounds consistent with what we have now, but to be honest I don't have the
> >background to agree or disagree. Could you explain (again) why does filename
> >have to reflect its option?
> 
> Because in a certain way, having the align option or not results in a
> different file (at least in a different behavior). While one could argue
> that in this case, it does not make that much of a difference, one can
> easily conceive scenarios in which it does (for instance, for blkdebug it
> might be an inject-error option). The idea is that the filename reported for
> a BDS should result in the same BDS being constructed when used, or at least
> one which behaves in the same way.

Who's making use of the json file name?

>If you open a file named blkdebug::x.img
> with -o inject-error.event=read_aio, then just reporting blkdebug::x.img
> would not be correct, since opening a BDS using the filename blkdebug::x.img
> alone would not result in a BDS behaving the same. Therefore, we have to
> generate a JSON object as the filename which includes this option.
> 
> So while we could solve the problem in this case by ignoring the "align"
> option when constructing the filename for blkdebug, it would not work for
> the general case.
>