Re: 'make check-tcg' fails with an assert in qemu_plugin_vcpu_init_hook

["Alex Bennée" <alex.bennee@linaro.org>]

Peter Maydell <peter.maydell@linaro.org> writes:

> On Wed, 21 Jun 2023 at 09:05, Alex Bennée <alex.bennee@linaro.org> wrote:
>>
>>
>> Peter Maydell <peter.maydell@linaro.org> writes:
>>
>> > On Tue, 20 Jun 2023 at 17:56, Peter Maydell <peter.maydell@linaro.org> wrote:
>> >>
>> >> $ make -C build/x86 check-tcg
>> >> make: Entering directory '/mnt/nvmedisk/linaro/qemu-from-laptop/qemu/build/x86'
>> >> [...]
>> >>   TEST    munmap-pthread on arm
>> >> **
>> >> ERROR:../../plugins/core.c:221:qemu_plugin_vcpu_init_hook: assertion
>> >> failed: (success)
>> >> **
>> >> ERROR:../../accel/tcg/cpu-exec.c:1024:cpu_exec_setjmp: assertion
>> >> failed: (cpu == current_cpu)
>> >
>> > git bisect blames commit d7ee93e2435970:
>> >
>> >     cputlb: Restrict SavedIOTLB to system emulation
>> >
>> > I think that commit is not correct, because it means that
>> > the size of 'struct CPUState' and also the offset of fields
>> > like 'cpu_index' will be different for files which are
>> > compile-per-target-for-usermode and files which are
>> > compile-once-only. The assert happens here because the
>> > code which sets up cpu_index is build-once, but the code
>> > in qemu_plugin_vcpu_init_hook() which reads cpu_index is
>> > build-per-target and now they don't agree about where in
>> > the struct the field is...
>>
>> Hmm two things from that imply:
>>
>>   - I suspect the plugin core stuff could be build once (or maybe twice,
>>     system and user)
>
> It is already build-once, that's why it goes wrong...

I thought it was the other way around:

  specific_ss.add(when: 'CONFIG_PLUGIN', if_true: [files(
    'loader.c',
    'core.c',
    'api.c',
  ), declare_dependency(link_args: plugin_ldflags)])

but if we built it for linux-user and softmmu this could be fixed (until
the next breakage anyway). cpus-common.c is the common code that sets
this once.

>>   - we need to have some guard rails somehow to make sure things don't
>>     go out of sync
>
> We do, this is the poison.h stuff. CONFIG_USER_ONLY is a
> special case which we don't poison because there would be
> too much refactoring required...

I guess a great big honking comment at the top of CPUState telling
people not to do that or pushing softmmu and user specific bits of
CPUState into their own de-referenced structures.

>
> -- PMM


-- 
Alex Bennée
Virtualisation Tech Lead @ Linaro