Re: [dpdk-dev] [PATCH] mk: fix missing link of librte_vhost in shared, non-combined config

[Panu Matilainen <pmatilai@redhat.com>]

On 02/13/2015 11:28 AM, Thomas Monjalon wrote:
> 2015-02-13 09:27, Panu Matilainen:
>> On 02/12/2015 05:44 PM, Thomas Monjalon wrote:
>>> 2015-02-11 12:31, Gonzalez Monroy, Sergio:
>>>> From: Panu Matilainen [mailto:pmatilai@redhat.com]
>>>>> On 02/11/2015 12:51 PM, Gonzalez Monroy, Sergio wrote:
>>>>>> I think that vhost is being linked in the wrong place (plugins section).
>>>>>> The plugins only get linked when building static libraries.
>>>>>> I think the patch should also remove vhost from the plugins section.
>>>>>
>>>>> Right, so vhost isn't a pluggable driver in the sense that pmds are. I wont
>>>>> claim to be familiar with all this virt-related puzzle pieces :) I'll send an
>>>>> updated patch, I was just looking to fix build in my particular config and
>>>>> ignored the rest.
>>>>>
>>>>> On a related note, shouldn't librte_pmd_bond and librte_pmd_xenvirt be
>>>>> included in the plugins section along with all the other pmds?
>>>>>
>>>> Hi Panu,
>>>>
>>>> Good  question :)
>>>>
>>>> I did wonder the same thing not long ago.
>>>>
>>>> I think the reason is that (someone may correct me if I'm wrong) there
>>>> are specific unit tests for those pmds (testing extra API) that require
>>>> them to always be linked against.
>>>
>>> A library is considered as a plugin if there is no public API and it
>>> registers itself. That's the case of normal PMD.
>>> But bonding and Xen have some library parts with public API.
>>> It has been discussed and agreed for bonding but I'm not aware of the Xen case.
>>
>> Fair enough, thanks for the explanation.
>>
>> Just wondering about versioning of these things - currently all the PMDs
>> are versioned as well, which is slightly at odds with their expected
>> usage, dlopen()'ed items usually are not versioned because it makes the
>> files moving targets. But if a plugin can be an library too then it
>> clearly needs to be versioned as well.
>
> Not sure to understand your considerations.
> Plugins must be versioned because there can be some incompatibilities
> like mbuf rework.

Plugins are version-dependent obviously, but the issue is somewhat 
different from library versioning. Plugins are generally consumers of 
the versioned ABIs, whereas libraries are the providers.

>
>> I'm just thinking of typical packaging where the unversioned *.so
>> symlinks are in a -devel subpackage and the versioned libraries are in
>> the main runtime package. Plugins should be loadable by a stable
>> unversioned name always, for libraries the linker handles it behind the
>> scenes. So in packaging these things, plugin *.so links need to be
>> handled differently (placed into the main package) from others. Not
>> rocket science to filter by 'pmd' in the name, but a new twist anyway
>> and easy to get wrong.
>>
>> One possibility to make it all more obvious might be having a separate
>> directory for plugins, the mixed case ccould be handled by symlinks.
>
> I think I don't understand which use case you are trying to solve.
>

Its a usability/documentation issue more than a technical one. If plugin 
DSO's are versioned (like they currently are), then loading them via eg 
-d becomes cumbersome since you need to hunt down and provide the 
versioned name, eg "testpmd -d librte_pmd_pcap.so.1 [...]"

Like said above, it can be worked around by leaving the unversioned 
symlinks in place for plugins in runtime (library) packages, but that 
sort of voids the point of versioning. One possibility would be 
introducing a per-version plugin directory that would be used as the 
default path for dlopen() unless an absolute path is used.

	- Panu -