Re: [RFC PATCH] qapi: Incorrect attempt to fix building with MC146818RTC=n

[Markus Armbruster]

Paolo Bonzini <address@hidden> writes:

> On 21/01/20 06:49, Markus Armbruster wrote:
>> Paolo Bonzini <address@hidden> writes:
>> 
>>> On 13/01/20 15:01, Markus Armbruster wrote:
>>>> Philippe Mathieu-Daudé <address@hidden> writes:
>>>>
>>>>> When configured with --without-default-devices and setting
>>>>> MC146818RTC=n, the build fails:
>>>>>
>>>>>     LINK    x86_64-softmmu/qemu-system-x86_64
>>>>>   /usr/bin/ld: qapi/qapi-commands-misc-target.o: in function 
>>>>> `qmp_marshal_rtc_reset_reinjection':
>>>>>   qapi/qapi-commands-misc-target.c:46: undefined reference to 
>>>>> `qmp_rtc_reset_reinjection'
>>>>>   /usr/bin/ld: qapi/qapi-commands-misc-target.c:46: undefined reference 
>>>>> to `qmp_rtc_reset_reinjection'
>>>>>   collect2: error: ld returned 1 exit status
>>>>>   make[1]: *** [Makefile:206: qemu-system-x86_64] Error 1
>>>>>   make: *** [Makefile:483: x86_64-softmmu/all] Error 2
>>>>>
>>>>> This patch tries to fix this, but this is incorrect because QAPI
>>>>> scripts only provide TARGET definitions, so with MC146818RTC=y we
>>>>> get:
>>>>>
>>>>>   hw/rtc/mc146818rtc.c:113:6: error: no previous prototype for 
>>>>> ‘qmp_rtc_reset_reinjection’ [-Werror=missing-prototypes]
>>>>>     113 | void qmp_rtc_reset_reinjection(Error **errp)
>>>>>         |      ^~~~~~~~~~~~~~~~~~~~~~~~~
>>>>>   cc1: all warnings being treated as errors
>>>>>   make[1]: *** [rules.mak:69: hw/rtc/mc146818rtc.o] Error 1
>>>>>
>>>>> Any idea? :)
>>>>>
>>>>> Signed-off-by: Philippe Mathieu-Daudé <address@hidden>
>>>>> ---
>>>>>  qapi/misc-target.json | 2 +-
>>>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>>>
>>>>> diff --git a/qapi/misc-target.json b/qapi/misc-target.json
>>>>> index a00fd821eb..8e49c113d1 100644
>>>>> --- a/qapi/misc-target.json
>>>>> +++ b/qapi/misc-target.json
>>>>> @@ -41,7 +41,7 @@
>>>>>  #
>>>>>  ##
>>>>>  { 'command': 'rtc-reset-reinjection',
>>>>> -  'if': 'defined(TARGET_I386)' }
>>>>> +  'if': 'defined(TARGET_I386) && defined(CONFIG_MC146818RTC)' }
>>>>>  
>>>>>  
>>>>>  ##
>>>>
>>>> The generated qapi-commands-misc-target.h duly has
>>>>
>>>>     #if defined(TARGET_I386) && defined(CONFIG_MC146818RTC)
>>>>     void qmp_rtc_reset_reinjection(Error **errp);
>>>>     void qmp_marshal_rtc_reset_reinjection(QDict *args, QObject **ret, 
>>>> Error **errp);
>>>>     #endif /* defined(TARGET_I386) && defined(CONFIG_MC146818RTC) */
>>>>
>>>> mc146818rtc.c includes it.  But since it doesn't include
>>>> config-devices.h, CONFIG_MC146818RTC remains undefined, and the
>>>> prototype gets suppressed.
>>>>
>>>> Crude fix: make mc146818rtc.c #include "config-devices.h".
>>>
>>> Can we modify the code generator to leave out the #if from the header,
>>> and only include it in the .c file?  An extra prototype is harmless.
>> 
>> Is *everything* we generate into headers just as harmless?
>
> It should be, since it's just the C version of some JSON.  The only
> problematic thing could be different definitions of the same command for
> multiple targets, and I think we want to avoid that anyway.
>
> To see it a different way, these are the "C bindings" to QMP, just that
> the implementation is an in-process call rather than RPC.  If the QAPI
> code generator was also able to generate Python bindings and the like,
> they would have to be the same for all QEMU binaries, wouldn't they?

Ommitting the kind of #if we've been discussing is relatively harmless:

    #if defined(TARGET_I386)
    void qmp_rtc_reset_reinjection(Error **errp);
    void qmp_marshal_rtc_reset_reinjection(QDict *args, QObject **ret, Error 
**errp);
    #endif /* defined(TARGET_I386) */

But what about this one, in qapi-types-block-core.h:

    typedef enum BlockdevDriver {
        BLOCKDEV_DRIVER_BLKDEBUG,
        [...]
    #if defined(CONFIG_REPLICATION)
        BLOCKDEV_DRIVER_REPLICATION,
    #endif /* defined(CONFIG_REPLICATION) */
        [...]
        BLOCKDEV_DRIVER__MAX,
    } BlockdevDriver;

If I omit it in the header, I then have to omit it in
qapi-types-block-core.c's

    const QEnumLookup BlockdevDriver_lookup = {
        .array = (const char *const[]) {
            [BLOCKDEV_DRIVER_BLKDEBUG] = "blkdebug",
            [...]
    #if defined(CONFIG_REPLICATION)
            [BLOCKDEV_DRIVER_REPLICATION] = "replication",
    #endif /* defined(CONFIG_REPLICATION) */
            [...]
        },
        .size = BLOCKDEV_DRIVER__MAX
    };

and God knows what else.  But I must not omit it in qapi-introspect.c's

        QLIT_QDICT(((QLitDictEntry[]) {
            { "meta-type", QLIT_QSTR("enum"), },
            { "name", QLIT_QSTR("245"), },
            { "values", QLIT_QLIST(((QLitObject[]) {
                QLIT_QSTR("blkdebug"),
                [...]
    #if defined(CONFIG_REPLICATION)
                QLIT_QSTR("replication"),
    #endif /* defined(CONFIG_REPLICATION) */
                [...]
                {}
            })), },
            {}
        })),

because that would defeat introspection.

I smell a swamp.

I'd rather not complicate the generator to support not including a
header I feel we *should* include.  #ifdef CONFIG_FOO can occur not just
in QAPI-generated code, and neglecting to include the relevant header
can cause *nasty* problems not just in QAPI-generated code.  Like
inconsistent struct definitions in separate compilation units.  Been
there, debugged that, wasn't fun, do not want to go there again.