Re: [PULL 00/30] ppc-for-5.2 queue 20200904

[Philippe Mathieu-Daudé]

On 9/7/20 7:26 PM, Laurent Vivier wrote:
> On 07/09/2020 18:29, Laurent Vivier wrote:
>> On 07/09/2020 16:51, Cornelia Huck wrote:
>>> On Mon, 7 Sep 2020 16:31:24 +0200
>>> Laurent Vivier <lvivier@redhat.com> wrote:
>>>
>>>> On 07/09/2020 16:05, Philippe Mathieu-Daudé wrote:
>>>>> Hi Thiago,
>>>>>
>>>>> On 9/7/20 3:29 PM, Laurent Vivier wrote:  
>>>>>> On 07/09/2020 04:38, David Gibson wrote:  
>>>>>>> On Sun, Sep 06, 2020 at 04:20:10PM +0100, Peter Maydell wrote:  
>>>
>>>>>>>> The 'check-tcg' tests for the linux-user static build also
>>>>>>>> failed on an s390x test:
>>>>>>>>
>>>>>>>>   CHECK   debian-s390x-cross
>>>>>>>>   BUILD   s390x-linux-user guest-tests with docker 
>>>>>>>> qemu/debian-s390x-cross
>>>>>>>>   RUN     tests for s390x
>>>>>>>>   TEST    threadcount on s390x
>>>>>>>> Unhandled trap: 0x10003  
>>>>>>
>>>>>> This is EXCP_HALTED (include/exec/cpu-all.h)
>>>>>>
>>>>>> The message error comes from cpu_loop() in linux-user/s390x/cpu_loop.c.
>>>>>>
>>>>>> The trap can only come from accel/tcg/cpu-exec.c
>>>>>>
>>>>>>     679 int cpu_exec(CPUState *cpu)
>>>>>>     680 {
>>>>>> ...
>>>>>>     688     if (cpu_handle_halt(cpu)) {
>>>>>>     689         return EXCP_HALTED;
>>>>>>     690     }
>>>>>>
>>>>>> and
>>>>>>
>>>>>>     428 static inline bool cpu_handle_halt(CPUState *cpu)
>>>>>>     429 {
>>>>>>     430     if (cpu->halted) {
>>>>>> ...
>>>>>>     441         if (!cpu_has_work(cpu)) {
>>>>>>     442             return true;
>>>>>>     443         }
>>>>>>
>>>>>> and
>>>>>>
>>>>>>      58 static bool s390_cpu_has_work(CPUState *cs)
>>>>>>      59 {
>>>>>>      60     S390CPU *cpu = S390_CPU(cs);
>>>>>>      61
>>>>>>      62     /* STOPPED cpus can never wake up */
>>>>>>      63     if (s390_cpu_get_state(cpu) != S390_CPU_STATE_LOAD &&
>>>>>>      64         s390_cpu_get_state(cpu) != S390_CPU_STATE_OPERATING) {
>>>>>>      65         return false;
>>>>>>      66     }
>>>>>>      67
>>>>>>      68     if (!(cs->interrupt_request & CPU_INTERRUPT_HARD)) {
>>>>>>      69         return false;
>>>>>>      70     }
>>>>>>      71
>>>>>>      72     return s390_cpu_has_int(cpu);
>>>>>>      73 }
>>>>>>
>>>>>> and in target/s390x/cpu.h:
>>>>>>
>>>>>>     772 #ifndef CONFIG_USER_ONLY
>>>>>>     773 unsigned int s390_cpu_set_state(uint8_t cpu_state, S390CPU *cpu);
>>>>>>     774 #else
>>>>>>     775 static inline unsigned int s390_cpu_set_state(uint8_t cpu_state,
>>>>>> S390CPU *cpu)
>>>>>>     776 {
>>>>>>     777     return 0;
>>>>>>     778 }
>>>>>>     779 #endif /* CONFIG_USER_ONLY */
>>>>>>     780 static inline uint8_t s390_cpu_get_state(S390CPU *cpu)
>>>>>>     781 {
>>>>>>     782     return cpu->env.cpu_state;
>>>>>>     783 }
>>>>>>
>>>>>> As cpu_state is never set, perhaps in case of linux-user it should
>>>>>> always return S390_CPU_STATE_OPERATING?
>>>
>>> Possibly, we should not have any state handling for linux-user.
>>>
>>
>> I did that, but now 390_cpu_has_work() is false because
>> CPU_INTERRUPT_HARD is not set in cs->interrupt_request.
>>
>> I think we should not enter in cpu_loop() with halted set to 1.
>>
>> Before the patch of this series, s390_cpu_reset() is called twice, and
>> on the second call, halted is already 0.
>>
>> With start_powered_off set to true in initfn, on the first reset
>> "halted" is 0 and on the second it is 1 (because it has been copied from
>> start_powered_off) and so cpu_loop() starts with halted set to 1 and fails.
> 
> What is happening:
> 
> [without start_powered_off]
> 
>   1- halted is set to 1 in s390x_cpu_initfn()
>   2- halted is set to 0 in s390x_cpu_reset() by parent_reset()
>      (cpu_common_reset()
>   3- cpu_loop() is always entered with halted set to 0
> 
> [with start_powered_off]
> 
>   1- halted is set to start_powered_off (1) in s390x_cpu_reset() by
>      parent_reset() (cpu_common_reset()
>   2- cpu_loop() is always entered with halted set to 1
> 
> So in the first case, cpu_loop() is always started with halted set to 0
> and in the second case with halted set to 1.
> 
> And I think, with linux-user, it should never be started with halted set
> to 1.
> 
> We can't add a "#ifdef CONFIG_USER_ONLY" in hw/core/cpu.c to set halted
> to 0 because it is in the common files, but we can do:
> 
> diff --git a/target/s390x/cpu.c b/target/s390x/cpu.c
> index 73d7d6007e8e..749cd548f0f3 100644
> --- a/target/s390x/cpu.c
> +++ b/target/s390x/cpu.c
> @@ -291,9 +291,9 @@ static void s390_cpu_initfn(Object *obj)
>      S390CPU *cpu = S390_CPU(obj);
> 
>      cpu_set_cpustate_pointers(cpu);
> -    cs->start_powered_off = true;
>      cs->exception_index = EXCP_HLT;
>  #if !defined(CONFIG_USER_ONLY)
> +    cs->start_powered_off = true;
>      object_property_add(obj, "crash-information", "GuestPanicInformation",
>                          s390_cpu_get_crash_info_qom, NULL, NULL, NULL);
>      cpu->env.tod_timer =

This looks like the correct fix indeed :)
(Maybe worth adding a comment around).

Thanks for investigating!

> 
> Thanks,
> Laurent