Re: [XEN PATCH v7 49/51] build: adding out-of-tree support to the xen build

[Jan Beulich <jbeulich@xxxxxxxx>]

On 18.10.2021 13:20, Juergen Gross wrote:
> On 18.10.21 13:07, Jan Beulich wrote:
>> On 18.10.2021 12:40, Juergen Gross wrote:
>>> On 18.10.21 12:36, Jan Beulich wrote:
>>>> On 18.10.2021 12:28, Juergen Gross wrote:
>>>>> On 18.10.21 11:51, Anthony PERARD wrote:
>>>>>> On Mon, Oct 18, 2021 at 11:02:20AM +0200, Jan Beulich wrote:
>>>>>>> On 15.10.2021 18:58, Anthony PERARD wrote:
>>>>>>>> I have to think about that. I've made some further progress in order to
>>>>>>>> be able to build the Xen pvhshim without a link farm and notice that
>>>>>>>> nearly every source file needs to use "$(srctree)/$(src)"
>>>>>>>
>>>>>>> Oh, now I'm curious as to the why here. I thought use of $(srctree)
>>>>>>> ought to be the exception.
>>>>>>
>>>>>> In Linux, the use of $(srctree) is indeed the exception. This is because
>>>>>> we have VPATH=$(srctree), so when `make` look for a prerequisite or a
>>>>>> target it will look first in the current directory and then in
>>>>>> $(srctree). That works fine as long as the source tree only have sources
>>>>>> and no built files.
>>>>>>
>>>>>> But if we want to be able to build the pv-shim without the linkfarm and
>>>>>> thus using out-of-tree build, we are going to need the ability to build
>>>>>> from a non-clean source tree. I don't think another way is possible.
>>>>>
>>>>> Is there any reason (apart from historical ones) to build the hypervisor
>>>>> in $(srctree)?
>>>>>
>>>>> I could see several advantages to build it in another directory as soon
>>>>> as the build system has this capability:
>>>>>
>>>>> - possibility to have a simple build target for building multiple archs
>>>>>      (assuming the cross-tools are available), leading to probably less
>>>>>      problems with breaking the build of "the other" architecture we are
>>>>>      normally not working with (and in future with e.g. Risc-V being added
>>>>>      this will be even more important)
>>>>>
>>>>> - possibility to have a debug and a non-debug build in parallel (in fact
>>>>>      at least at SUSE we are working around that by building those with an
>>>>>      intermediate "make clean" for being able to package both variants)
>>>>>
>>>>> - make clean for the hypervisor part would be just a "rm -r"
>>>>
>>>> I fully agree, yet ...
>>>>
>>>>> Yes, this would require us (the developers) to maybe change some habits,
>>>>> but I think this would be better than working around the issues by
>>>>> adding $(srctree) all over the build system.
>>>>
>>>> ... developers' habits would only be my second concern here (and if that
>>>> had been the only one, then I would not see this as a reason speaking
>>>> against the change, but as said I've never been building from the root,
>>>> and I've also been building sort of out-of-tree all the time). Yet while
>>>> writing this reply I came to realize that my primary concern was wrong:
>>>> People would not need to adjust their spec files (or alike), at least
>>>> not as long as they consume only files living under dist/.
>>>>
>>>> So, Anthony - thoughts about making the default in-tree Xen build
>>>> actually build into, say, build/xen/?
>>>
>>> Or maybe even build-<arch>[-debug]/xen/?
>>
>> I'd be okay with build-<arch>, but things would become questionable imo
>> when considering further elements recorded in .config: Where would you
>> draw the line?
> 
> Okay, this is a valid question. What about an environment variable which
> can be used to determine the build directory (or a suffix of the build
> directory)?

That would be fine with me, but as said I'm not building from the root
dir anyway, so people affected by a possible change here may be more
qualified to give input. Otoh if people need to adjust their stuff
anyway, they can as well switch to an out-of-tree Xen build right away.
And anyone building in-tree won't expect to be able to build multiple
flavors in parallel.

Jan