Re: [Qemu-devel] [PATCH for-4.0 v7 01/27] qapi: make sure osdep.h is included in type headers

[Markus Armbruster]

Marc-André Lureau <address@hidden> writes:

> Hi
>
> On Mon, Dec 10, 2018 at 5:28 PM Markus Armbruster <address@hidden> wrote:
>>
>> Marc-André Lureau <address@hidden> writes:
>>
>> > Hi
>> >
>> > On Mon, Dec 10, 2018 at 1:52 PM Markus Armbruster <address@hidden> wrote:
>> >>
>> >> Marc-André Lureau <address@hidden> writes:
>> >>
>> >> > Now that the schema can be configured, it is crucial that all types
>> >> > are configured the same. Make sure config-host.h is included, by
>> >> > checking osdep.h inclusion. The build-sys tracks the dependency and
>> >> > rebuilds the types if the configuration changed.
>> >> >
>> >> > Signed-off-by: Marc-André Lureau <address@hidden>
>> >> > ---
>> >> >  scripts/qapi/types.py | 3 +++
>> >> >  1 file changed, 3 insertions(+)
>> >> >
>> >> > diff --git a/scripts/qapi/types.py b/scripts/qapi/types.py
>> >> > index fd7808103c..3bb9cb6d47 100644
>> >> > --- a/scripts/qapi/types.py
>> >> > +++ b/scripts/qapi/types.py
>> >> > @@ -201,6 +201,9 @@ class 
>> >> > QAPISchemaGenTypeVisitor(QAPISchemaModularCVisitor):
>> >> >  ''',
>> >> >                                        types=types, visit=visit))
>> >> >          self._genh.preamble_add(mcgen('''
>> >> > +#ifndef QEMU_OSDEP_H
>> >> > +#error Please include osdep.h first!
>> >> > +#endif
>> >> >  #include "qapi/qapi-builtin-types.h"
>> >> >  '''))
>> >>
>> >> I understand why you propose this patch.  The QAPI-generated headers use
>> >> #ifdef CONFIG_FOO.  The configuration header "qemu/osdep.h" must be
>> >> consistently included before the generated headers, or else you end up
>> >> with nasty bugs, such as the same enum having different values in
>> >> different .o, or the same struct having a different layout.
>> >>
>> >> But this applies to *all* headers that use #ifdef.  Why check it here,
>> >> but not there?  What makes the QAPI-generated headers special?
>> >>
>> >
>> > It's the discussion about #if in headers (and enums in particular)
>> > that started this. We want to make sure all compilation units share
>> > the same data structure/ABI. I proposed to include osdep.h in qapi
>> > headers, but that was rejected.
>> > The warning is a different approach. I agree it could apply to all
>> > headers. Do you think I should update all headers as well?
>>
>> That would replace the rule 'all .c files must include "qemu/osdep.h"
>> first' by 'all .h files must check "qemu/osdep.h" has been included
>> already', which is not an improvement.
>
> That would be an improvement since headers may also be impacted by osdep.h
>
> Alternatively, why not use -include ?

Requiring .c to include the configuration header first follows
established Autoconf practice.  The "first" part covers all headers.

I've seen plenty of autoconfiscated code, yet none where the headers
double-check the configuration header has been included already.  Such a
check is neither sufficient nor really necessary.  Checking the .c is
both.

I have no strong feelings about using -include instead.  I'd prefer to
avoid the difference to what other projects do, though.

Circling back to the issue that made you propose this patch.

I think I'm (belatedly) coming around to your initial view that the use
of conditionals in generated QAPI code is safe enough.

I shouldn't be worried about #if defined(CONFIG_HOST_THING) where
CONFIG_HOST_THING belongs to config-host.h.  That's always pulled in via
qemu/osdep.h.

I also shouldn't be worried about #if defined(CONFIG_TARGET_THING) where
CONFIG_TARGET_THING belongs to config-target.h.  qemu/osdep.h pulls in
either that or exec/poison.h.

I even shouldn't be worried about #if defined(RANDOM_MACRO) as long as
qemu/osdep.h pulls in its owner or poisons it.  I might legitimately be
worried about the "as long as" part.

We could perhaps extract the ifconds and look for macros that belong
neither to config-host.h nor config-target.h.  But I'm not sure it's
worth the bother.