Re: Getting rid of (many) dynamic link creations in the xen build

[Andrew Cooper <andrew.cooper3@xxxxxxxxxx>]

On 15/10/2020 11:41, Jürgen Groß wrote:
> On 15.10.20 12:09, Jan Beulich wrote:
>> On 15.10.2020 09:58, Jürgen Groß wrote:
>>> After a short discussion on IRC yesterday I promised to send a mail
>>> how I think we could get rid of creating dynamic links especially
>>> for header files in the Xen build process.
>>>
>>> This will require some restructuring, the amount will depend on the
>>> selected way to proceed:
>>>
>>> - avoid links completely, requires more restructuring
>>> - avoid only dynamically created links, i.e. allowing some static
>>>     links which are committed to git
>>
>> While I like the latter better, I'd like to point out that not all
>> file systems support symlinks, and hence the repo then couldn't be
>> stored on (or the tarball expanded onto) such a file system. Note
>> that this may be just for viewing purposes - I do this typically at
>> home -, i.e. there's no resulting limitation from the build process
>> needing symlinks. Similarly, once we fully support out of tree
>> builds, there wouldn't be any restriction from this as long as just
>> the build tree is placed on a capable file system.
>>
>> As a result I'd like to propose variant 2´: Reduce the number of
>> dynamically created symlinks to a minimum. This said, I have to
>> admit that I haven't really understood yet why symlinks are bad.
>> They exist for exactly such purposes, I would think.
>
> Not the symlinks as such, but the dynamically created ones seem to be
> a problem, as we stumble upon those again and again.

We have multiple build system bugs every release to do with dynamically
generated symlinks.  Given that symlinks aren't a hard requirement, this
is a massive price to pay, and time which could be better spent doing
other other things.

Also, they prohibit the ability to build from a read-only source dir.

The asm symlink in particular prevents any attempt to do concurrent
builds of xen.  In some future, I'd love to be able to do concurrent
out-of-tree builds of Xen on different architectures, because elapsed
time to do this is one limiting factor of mine on pre-push sanity checks.

Personally, I'd prefer option 1 in the long run, but I've got no
problems with achieving option 2 as an intermediate goal.

>
>>
>>> The difference between both variants is affecting the public headers
>>> in xen/include/public/: avoiding even static links would require to
>>> add another directory or to move those headers to another place in the
>>> tree (either use xen/include/public/xen/, or some other path */xen),
>>> leading to the need to change all #include statements in the hypervisor
>>> using <public/...> today.
>>>
>>> The need for the path to have "xen/" is due to the Xen library headers
>>> (which are installed on user's machines) are including the public
>>> hypervisor headers via "#include <xen/...>" and we can't change that
>>> scheme. A static link can avoid this problem via a different path, but
>>> without any link we can't do that.
>>>
>>> Apart from that decision, lets look which links are created today for
>>> accessing the header files (I'll assume my series putting the library
>>> headers to tools/include will be taken, so those links being created
>>> in staging today are not mentioned) and what can be done to avoid them:
>>>
>>> - xen/include/asm -> xen/include/asm-<arch>:
>>>     Move all headers from xen/include/asm-<arch> to
>>>     xen/arch/<arch>/include/asm and add that path via "-I" flag to
>>> CFLAGS.
>>>     This has the other nice advantages that most architecture specific
>>>     files are now in xen/arch (apart from the public headers) and
>>> that we
>>>     can even add generic fallback headers in xen/include/asm in case an
>>>     arch doesn't need a specific header file.
>>
>> Iirc Andrew suggested years ago that we follow Linux in this regard
>> (and XTF already does). My only concern here is the churn this will
>> cause for backports.
>
> Changing a directory name in a patch isn't that hard, IMO.

Also git (if you throw it the correct runes) can cope with this
automatically.

>>> - xen/arch/<arch>/efi/*.[ch] -> xen/common/efi/*.[ch]:
>>>     Use vpath for the *.c files and the "-I" flag for adding
>>> common/efi to
>>>     the include path in the xen/arch/<arch>/efi/Makefile.
>>
>> Fine with me.

Something which has been irritating me for years is that cscope doesn't
tollerate the efi symlinks.  This would be a great solution.

~Andrew