On 5/8/2026 1:29 PM, Richard Henderson wrote:
> On 5/5/26 17:48, Pierrick Bouvier wrote:
>> For now, we expect only one target to be available at runtime. This will
>> change with the single-binary and we'll detect which one to use
>> dynamically.
>>
>> Reviewed-by: Marc-André Lureau <[email protected]>
>> Signed-off-by: Pierrick Bouvier <[email protected]>
>> ---
>>   include/qemu/target-info-qom.h |  2 ++
>>   system/vl.c                    |  2 ++
>>   target-info-qom.c              | 16 ++++++++++++++++
>>   3 files changed, 20 insertions(+)
>>
>> diff --git a/include/qemu/target-info-qom.h b/include/qemu/target-
>> info-qom.h
>> index 585995c7ad0..91be415ed33 100644
>> --- a/include/qemu/target-info-qom.h
>> +++ b/include/qemu/target-info-qom.h
>> @@ -25,4 +25,6 @@ typedef struct TargetInfoQomClass {
>>     OBJECT_DECLARE_TYPE(TargetInfoQom, TargetInfoQomClass, TARGET_INFO)
>>   +void target_info_qom_set_target(void);
>> +
>>   #endif /* QEMU_TARGET_INFO_QOM_H */
>> diff --git a/system/vl.c b/system/vl.c
>> index 2b6739271ba..77b2b4e673d 100644
>> --- a/system/vl.c
>> +++ b/system/vl.c
>> @@ -28,6 +28,7 @@
>>   #include "qemu/units.h"
>>   #include "qemu/module.h"
>>   #include "qemu/target-info.h"
>> +#include "qemu/target-info-qom.h"
>>   #include "exec/cpu-common.h"
>>   #include "exec/page-vary.h"
>>   #include "hw/core/qdev-properties.h"
>> @@ -2891,6 +2892,7 @@ void qemu_init(int argc, char **argv)
>>       os_setup_limits();
>>         module_call_init(MODULE_INIT_TARGET_INFO);
>> +    target_info_qom_set_target();
>>         module_init_info(qemu_modinfo);
>>       module_allow_arch(target_name());
>> diff --git a/target-info-qom.c b/target-info-qom.c
>> index ba2c7923760..9d1f50ffcab 100644
>> --- a/target-info-qom.c
>> +++ b/target-info-qom.c
>> @@ -36,3 +36,19 @@ static const TypeInfo target_info_parent_type = {
>>   };
>>     DEFINE_TARGET_INFO_TYPE(target_info_parent_type)
>> +
>> +static const TargetInfo *target_info_ptr;
>> +
>> +void target_info_qom_set_target(void)
>> +{
>> +    g_autoptr(GSList) targets =
>> object_class_get_list(TYPE_TARGET_INFO, false);
>> +
>> +    size_t num_found = g_slist_length(targets);
>> +    if (num_found != 1) {
>> +        error_setg(&error_fatal, num_found == 0 ?
>> +                                 "no target-info is available" :
>> +                                 "more than one target-info is
>> available");
>> +    }
>> +
>> +    target_info_ptr = TARGET_INFO_CLASS(targets->data)->target_info;
>> +}
> 
> Reviewed-by: Richard Henderson <[email protected]>
> 
> For future improvement, I suggest structuring the code more like page-
> vary-common.c, where exactly one compilation unit sees a non-const
> "TargetInfo target_info", all others see a "const TargetInfo
> target_info", and target_info() the function is eliminated.
> 
> One could even plausibly unify page-vary with target-info, rather than
> duplicating some of the page_bits information between the two structures.
> 
> Most pieces of target_info are really only referenced at startup, but
> target_big_endian() and target_long_bits() are the big exceptions to
> that, and it would be really nice for them not to require an out-of-line
> function call.
>

The only point we are still discussing with Philippe is how we'll model
having several targets, especially at the startup, where we create
machines and types. Once you're in a vcpu thread, things are easier as
there is only one context possible.

It does not mean we can't eliminate the function now, and using indirect
struct access instead, but I think we'll need to have something else to
answer the startup phase. And we don't have a clear answer yet.
 >
> r~

Regards,
Pierrick

Reply via email to