Re: [Qemu-devel] [PATCH] Modify qemu-img can not create local filename contain ":"

[Eric Blake]

On 03/03/2014 10:47 AM, Max Reitz wrote:

> 
> Currently, a protocol prefix is only defined to end with a colon, not
> with ":/" or "://". There are already protocol block drivers which do
> not require a slash after the colon such as blkdebug or blkverify and I
> deem it rather impossible to redefine their filename format now (in
> order to make them use ":/").

It should ALWAYS be possible to open a file via absolute path, or via
'./file:with:colon'.  That is, our check for a protocol should be any
prefix that contains a ':' prior to a '/'.

> 
> Thus, I do not think it will be this easy. I am in fact not even sure
> whether it is possible at all (automagically doing the right thing) –
> currently, a colon simpy is the separator between protocol and filename.
> Just using the "file" protocol per default for the whole filename if
> there is no protocol with a name equal to the pre-colon part is probably
> not what we want, since the user may actually be referring to some
> protocol that the qemu version he/she is using does not support (yet),
> in which case creating a file with a name including the pre-colon part
> (the protocol name) is most probably not the right thing to do.

If the user wants a file name that contains a colon, use either an
absolute path name, or prefix the relative name with './'.  Either way
should cause a '/' to occur before the first ':', and since a protocol
consists only of a prefix prior to ':' without '/', it then becomes
obvious that a file name is intended.

> 
> Henceforth, in my opinion, we either have to ask the user in that case
> or we introduce some new option which disables protocol prefixes. I
> think the easiest way to do the latter is to introduce a
> bdrv_parse_filename() function for the "file" protocol drivers which
> remove a "file:" prefix if given. Then, the user could just specify
> "file:foo:bar.img" to reference a file named "foo:bar.img".

Yes, this would also make sense, if you can't live with './foo:bar.img'.

> 
> Currently, the behavior is that such a prefix will be interpreted
> correctly (the "file" protocol is selected) but it is not removed. Thus,
> "file:foo:bar.img" will actually reference a file named
> "file:foo:bar.img". One could argue that removing the prefix therefore
> breaks current behavior, but I sincerely hope nobody has relied on that
> behavior so far.

I think the fact that the file: prefix is NOT removed before hitting the
filesystem is a bug worth fixing - it should be fairly obvious that the
relative name in the filesystem should not include the 'file:' prefix,
but that 'file:' was the protocol that forced interpretation of the rest
of the string as a local filename.

-- 
Eric Blake   eblake redhat com    +1-919-301-3266
Libvirt virtualization library http://libvirt.org

signature.asc
Description: OpenPGP digital signature