Re: [PATCH v2 1/3] target/arm: Restrict v8M IDAU to TCG

[Philippe Mathieu-Daudé]

On 3/10/21 12:46 PM, Claudio Fontana wrote:
> On 3/9/21 3:18 PM, Philippe Mathieu-Daudé wrote:
>> On 3/9/21 2:41 PM, Claudio Fontana wrote:
>>> On 2/21/21 11:26 PM, Philippe Mathieu-Daudé wrote:
>>>> IDAU is specific to M-profile. KVM only supports A-profile.
>>>> Restrict this interface to TCG, as it is pointless (and
>>>> confusing) on a KVM-only build.
>>>>
>>>> Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
>>>> Reviewed-by: Peter Maydell <peter.maydell@linaro.org>
>>>> Signed-off-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
>>>
>>>
>>> This one breaks the KVM tests hard though (most of them).
>>>
>>> I will try to figure out why.
>>>
>>> Ciao,
>>>
>>> Claudio
>>>
>>>
>>>> ---
>>>>  target/arm/cpu.c     | 7 -------
>>>>  target/arm/cpu_tcg.c | 8 ++++++++
>>>>  2 files changed, 8 insertions(+), 7 deletions(-)
>>>>
>>>> diff --git a/target/arm/cpu.c b/target/arm/cpu.c
>>>> index b8bc89e71fc..a772fd4926f 100644
>>>> --- a/target/arm/cpu.c
>>>> +++ b/target/arm/cpu.c
>>>> @@ -2380,12 +2380,6 @@ static const TypeInfo arm_cpu_type_info = {
>>>>      .class_init = arm_cpu_class_init,
>>>>  };
>>>>  
>>>> -static const TypeInfo idau_interface_type_info = {
>>>> -    .name = TYPE_IDAU_INTERFACE,
>>>> -    .parent = TYPE_INTERFACE,
>>
>> Hmm this is an interface...
>>
>> Is a CPU/machine trying to resolve it?
> 
> Well, this fails horribly at any qemu-system-aarch64 startup for the kvm-only 
> build:
> 
> in my view we cannot remove the idau interface until we have removed all the 
> TCG-only boards fronm the build.

Yes, this is a similar bug to the one fixed by commit 8d0bceba24c
("hw/nvram: Always register FW_CFG_DATA_GENERATOR_INTERFACE").

> 
> When calling qemu_init(), and we get into select_machine(),
> 
> the object_class_get_list() tries to initialize all machine types.
> 
> When it does that, it tries to initialize the IDAU interface, and fails.
> 
> #0  0x0000ffffb9e51828 in raise () at /lib64/libc.so.6
> #1  0x0000ffffb9e52e4c in abort () at /lib64/libc.so.6
> #2  0x0000aaaae042a484 in type_initialize (ti=0xaaaaf0cb37c0) at 
> ../qom/object.c:333
> #3  0x0000aaaae042c06c in object_class_foreach_tramp (key=0xaaaaf0cb3940, 
> value=0xaaaaf0cb37c0, opaque=0xfffff9f2bac8)
>     at ../qom/object.c:1069
> #4  0x0000ffffbb3d4248 in g_hash_table_foreach () at 
> /usr/lib64/libglib-2.0.so.0
> #5  0x0000aaaae042c180 in object_class_foreach (fn=
>     0xaaaae042c324 <object_class_get_list_tramp>, 
> implements_type=0xaaaae089cc90 "machine", include_abstract=false, 
> opaque=0xfffff9f2bb10)
>     at ../qom/object.c:1091
> #6  0x0000aaaae042c3a8 in object_class_get_list 
> (implements_type=0xaaaae089cc90 "machine", include_abstract=false) at 
> ../qom/object.c:1148
> #7  0x0000aaaae03863d8 in select_machine () at ../softmmu/vl.c:1607
> #8  0x0000aaaae038ad74 in qemu_init (argc=15, argv=0xfffff9f2be08, 
> envp=0xfffff9f2be88) at ../softmmu/vl.c:3489
> #9  0x0000aaaadfdcf5a0 in main (argc=15, argv=0xfffff9f2be08, 
> envp=0xfffff9f2be88) at ../softmmu/main.c:49
> 
> 
> (gdb) frame 2
> #2  0x0000aaaae042a484 in type_initialize (ti=0xaaaaf0cb37c0) at 
> ../qom/object.c:333
> 333                     abort();
> (gdb) p ti[0]
> $1 = {name = 0xaaaaf0cb3940 "mps2tz", class_size = 408, instance_size = 
> 202224, instance_align = 0, class_init = 
>     0xaaaae0273408 <mps2tz_class_init>, class_base_init = 0x0, class_data = 
> 0x0, instance_init = 0x0, instance_post_init = 0x0, 
>   instance_finalize = 0x0, abstract = true, parent = 0xaaaaf0cb3960 
> "machine", parent_type = 0xaaaaf0cad860, class = 0xaaaaf0d0d830, 
>   num_interfaces = 1, interfaces = {{typename = 0xaaaaf0cb3980 
> "idau-interface"}, {typename = 0x0} <repeats 31 times>}}
> 
> 
> In my view we should revert this until all incompatible boards are disabled

My view is this is a QOM design problem. Others might hit the
same issue. It is hard to debug. It should be fixed upfront.

> In this case, the one failing is MPS2, so the offender is
> 
> devices/arm-softmmu.mak:CONFIG_MPS2=y
> 
> from the point of view of the kvm-only build.
> 
> What I'd suggest is (but I am open to alternatives):
> 
> * revert this one
> * complete my arm cleanup series, with now all tests passing
> * disable the non-KVM boards for KVM-only builds (basically your series)
> * apply the accelerator classes specializations to ARM

MPS2TZ uses a Cortex-M33 which is requires TCG.
The machine shouldn't be present if TCG is not built-in.

Previous attempt which you acked :)
"target/arm: Restrict ARMv7 M-profile cpus to TCG accel"
https://www.mail-archive.com/qemu-block@nongnu.org/msg79943.html