Re: [PULL 00/13] Testing, build system and misc patches

[Paolo Bonzini <pbonzini@redhat.com>]

On 06/09/21 11:51, Thomas Huth wrote:
> On 03/09/2021 18.49, Peter Maydell wrote:
>> On Fri, 3 Sept 2021 at 17:37, Alex Bennée <alex.bennee@linaro.org> wrote:
>>> Thomas Huth <thuth@redhat.com> writes:
>>>> On 03/09/2021 15.22, Peter Maydell wrote:
>>>>> This provokes a new warning from meson on a linux-static build:
>>>>> Run-time dependency appleframeworks found: NO (tried framework)
>>>>> Library rt found: YES
>>>>> Found pkg-config: /usr/bin/pkg-config (0.29.1)
>>>>> WARNING: Static library 'gbm' not found for dependency 'gbm', may not
>>>>> be statically linked
>>>>> Run-time dependency gbm found: YES 20.0.8
>>>>> Dependency libpng found: YES 1.6.34 (cached)
>>>>> Dependency libjpeg found: YES unknown (cached)
>>>>> If we're building statically and we can't find a static
>>>>> library then (a) we shouldn't print a WARNING and
>>>>> (b) we shouldn't then conclude that we've found gdm.
>>>>
>>>> Hmmm, no clue what's wrong here, since I basically did declare it like
>>>> all other libraries are declared, too (so this problem should have
>>>> shown up somewhere else already?)... Paolo, do you have any ideas
>>>> what's going on here?
>>>
>>> In attempting to replicate I found all the dynamic libs blow up:
>>
>>>    WARNING: Static library 'xkbcommon' not found for dependency 
>>> 'xkbcommon', may not be statically l
>>>    Run-time dependency xkbcommon found: YES 1.0.3
>>
>> I do vaguely recall complaining about new meson warnings for
>> static library detection in the past as well:
>> https://lore.kernel.org/qemu-devel/CAFEAcA8chPqS0keyGv0vBgNgacnMo95gA3LZDU2QfmteQ=4UZg@mail.gmail.com/ 
>>
>> https://lore.kernel.org/qemu-devel/CAFEAcA_-cNmt-sY3nqnGkpUqET86M6-82rf-Uv3QkwCR14kYsw@mail.gmail.com/ 
>>
>> https://lore.kernel.org/qemu-devel/CAFEAcA8xHxCGhh2hibsdCxZrYRRU+xcwVsa85O7KL9BsmW7ohw@mail.gmail.com/ 
>>
>>
>>> So is this a general problem with static libs. BTW I didn't catch this
>>> because I only build user with --static as I thought system --static was
>>> flakey anyway.
>>
>> I'm not doing a system build in this case... Looking at some of
>> those older threads, it looks like part of the answer is that
>> for dependencies that we don't need for linux-user mode we should
>> guard the test with some suitable if condition so we don't create
>> the dependency unless we're going to use it, eg the brlapi check
>> uses "if not get_option('brlapi').auto() or have_system", rbd
>> has a similar thing involving have_block, etc.
> 
> Ok, thanks, that seems to work, I'll change the patch accordingly.
> 
>> But I think there is an underlying meson bug here which that kind of
>> use of an if is merely working around: if we ask for a static library
>> it should not give us a dynamic library.
> 
> Agreed. Actually, when I run configure with "--static --disable-system" 
> on my laptop, I'm also getting some warnings:
> 
> WARNING: Static library 'z' not found for dependency 'zlib', may not be 
> statically linked
> Run-time dependency zlib found: YES 1.2.11
> Run-time dependency appleframeworks found: NO (tried framework)
> Library rt found: YES
> WARNING: Static library 'png16' not found for dependency 'libpng', may 
> not be statically linked
> WARNING: Static library 'z' not found for dependency 'libpng', may not 
> be statically linked
> 
> ... and linking then later fails while running "make".
> 
> Paolo, could the behavior of meson be changed to fail already the 
> configuration step in this case instead of only printing a warning?

The reason why this is just a warning is explained only in the code, and
it's this:

     # Library wasn't found, maybe we're looking in the wrong
     # places or the library will be provided with LDFLAGS or
     # LIBRARY_PATH from the environment (on macOS), and many
     # other edge cases that we can't account for.
     #
     # Add all -L paths and use it as -lfoo

In other words, Meson doesn't really know the library will be used for a
statically-linked binary (as opposed to just not wanting a shared
library for whatever reason).  So it looks for a .a file, and forces use
of the a static library by passing a path to that file.  If it cannot
find one, it warns.

Note that pre-Meson we didn't warn for --disable-system (correct) but we
did the wrong thing silently for --enable-system (just like now, except
without a warning).  So Meson's warning forces us to be a bit more verbose
to do only strictly necesary tests, as in:

   pam = not_found
   if not get_option('auth_pam').auto() or have_system
     pam = cc.find_library('pam', has_headers: ['security/pam_appl.h'],
                           required: get_option('auth_pam'),
                           kwargs: static_kwargs)
   endif

but it also catches incorrect setups on the user side and makes the "configure"
step a little faster with --disable-system.

FWIW, in the latest Meson version there's a shortcut for the above pattern,
since it is very common in QEMU; it can be rewritten as follows to avoid the
if/endif:

   pam = cc.find_library('pam', has_headers: ['security/pam_appl.h'],
                         required: get_option('auth_pam').disable_auto_if(not have_system)
                         kwargs: static_kwargs)

Paolo