Re: [Qemu-devel] [PATCH 4/4] build: don't build hardware objects with linux-user

[Laurent Vivier]

On 26/03/2019 10:00, Paolo Bonzini wrote:
> On 21/03/19 13:46, Laurent Vivier wrote:
>> +
>> +ifneq ($(CONFIG_USER_ONLY),y)
>>  #######################################################################
>>  # authz-obj-y is code used by both qemu system emulation and qemu-img
>>  
>> @@ -21,21 +27,11 @@ block-obj-$(CONFIG_REPLICATION) += replication.o
>>  
>>  block-obj-m = block/
> 
> Two remarks:
> 
> 1) Isn't CONFIG_USER_ONLY set to y even if tools are built?

Yes, you right. I didn't test combination "--enable-user --enable-tools
--disable-system"


> 
> 2) Notwithstanding the answer to (1), why is this part needed at all?
> The only things that have dependencies on crypto-obj-y, io-obj-y etc.
> are $(SOFTMMU_SUBDIR_RULES), the tools, and the tests that you are
> disabling already.  Therefore, why are they built at all for user-only
> builds?

The change was based on the comments in the file.

And for the crypto part we need it and this could explain:

commit fb37726db77b21f3731b90693d2c93ade1777528
Author: Daniel P. Berrange <address@hidden>
Date:   Wed Sep 2 10:57:27 2015 +0100

    crypto: move crypto objects out of libqemuutil.la

    Future patches will be adding more crypto related APIs which
    rely on QOM infrastructure. This creates a problem, because
    QOM relies on library constructors to register objects. When
    you have a file in a static .a library though which is only
    referenced by a constructor the linker is dumb and will drop
    that file when linking to the final executable :-( The only
    workaround for this is to link the .a library to the executable
    using the -Wl,--whole-archive flag, but this creates its own
    set of problems because QEMU is relying on lazy linking for
    libqemuutil.a. Using --whole-archive majorly increases the
    size of final executables as they now contain a bunch of
    object code they don't actually use.

    The least bad option is to thus not include the crypto objects
    in libqemuutil.la, and instead define a crypto-obj-y variable
    that is referenced directly by all the executables that need
    this code (tools + softmmu, but not qemu-ga). We avoid pulling
    entire of crypto-obj-y into the userspace emulators as that
    would force them to link to gnutls too, which is not required.



> 3) I don't really see the need or advantage in conditionalizing tests of
> libqemuutil.a, which is built even for linux-user builds (even though
> e.g. the coroutines code does not end up in the binary).  Since most of
> the tests you are removing are related to the block layer, perhaps you
> can add
> 
> CONFIG_BLOCK := $(call lor, $(CONFIG_SOFTMMU), $(call notempty,$(TOOLS))
> 
> and use it to conditionalize the testcases that depend on block-obj-y?
> 

I agree.

Thanks,
Laurent