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

[Philippe Mathieu-Daudé]

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:
>>> On Fri, 4 Sep 2020 at 04:47, David Gibson <david@gibson.dropbear.id.au> 
>>> wrote:
>>>>
>>>> The following changes since commit 
>>>> 67a7bfe560a1bba59efab085cb3430f45176d382:
>>>>
>>>>   Merge remote-tracking branch 
>>>> 'remotes/huth-gitlab/tags/pull-request-2020-09-03' into staging 
>>>> (2020-09-03 16:58:25 +0100)
>>>>
>>>> are available in the Git repository at:
>>>>
>>>>   git://github.com/dgibson/qemu.git tags/ppc-for-5.2-20200904
>>>>
>>>> for you to fetch changes up to b172606ecf29a140073f7787251a9d70ecb53b6e:
>>>>
>>>>   spapr_numa: move NVLink2 associativity handling to spapr_numa.c 
>>>> (2020-09-04 13:40:09 +1000)
>>>>
>>>> ----------------------------------------------------------------
>>>> ppc patch queue 2020-09-04
>>>>
>>>> Next pull request for qemu-5.2.  The biggest thing here is the
>>>> generalization of ARM's start-powered-off machine property to all
>>>> targets.  This can fix a number of odd little edge cases where KVM
>>>> could run vcpus before they were properly initialized.  This does
>>>> include changes to a number of files that aren't normally in my
>>>> purview.  There are suitable Acked-by lines and Peter requested this
>>>> come in via my tree, since the most pressing requirement for it is in
>>>> pseries machines with the POWER secure virtual machine facility.
>>>>
>>>> In addition we have:
>>>>  * The start of Daniel Barboza's rework and clean up of pseries
>>>>    machine NUMA handling
>>>>  * Correction to behaviour of the nvdimm= generic machine property on
>>>>    pseries
>>>>  * An optimization to the allocation of XIVE interrupts on KVM
>>>>  * Some fixes for confused behaviour with kernel_irqchip when both
>>>>    XICS and XIVE are in play
>>>>  * Add HIOMAP comamnd to pnv flash
>>>>  * Properly advertise the fact that spapr_vscsi doesn't handle
>>>>    hotplugged disks
>>>>  * Some assorted minor enhancements
>>>
>>> Hi -- this fails to build for Windows:
>>>
>>> ../../hw/ppc/spapr_numa.c: In function 'spapr_numa_fixup_cpu_dt':
>>> ../../hw/ppc/spapr_numa.c:77:5: error: unknown type name 'uint'
>>>      uint vcpu_assoc_size = NUMA_ASSOC_SIZE + 1;
>>>      ^
>>
>> Huh, that's weird.  My testing run was less thorough than I'd usually
>> do, because so many tests were broken on the master branch, but I was
>> pretty sure I did do successful mingw builds.
>>
>>> That should probably be using one of the standard C types.
>>
>> Done.
>>
>>> 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?
> 
> Something like:
> 
> diff --git a/target/s390x/cpu.h b/target/s390x/cpu.h
> index 035427521cec..8a8628fcdcc6 100644
> --- a/target/s390x/cpu.h
> +++ b/target/s390x/cpu.h
> @@ -771,16 +771,20 @@ int s390_assign_subch_ioeventfd(EventNotifier
> *notifier, uint32_t sch_id,
>                                  int vq, bool assign);
>  #ifndef CONFIG_USER_ONLY
>  unsigned int s390_cpu_set_state(uint8_t cpu_state, S390CPU *cpu);
> +static inline uint8_t s390_cpu_get_state(S390CPU *cpu)
> +{
> +    return cpu->env.cpu_state;
> +}
>  #else
>  static inline unsigned int s390_cpu_set_state(uint8_t cpu_state,
> S390CPU *cpu)
>  {
>      return 0;
>  }
> -#endif /* CONFIG_USER_ONLY */
>  static inline uint8_t s390_cpu_get_state(S390CPU *cpu)
>  {
> -    return cpu->env.cpu_state;
> +    return S390_CPU_STATE_OPERATING;
>  }
> +#endif /* CONFIG_USER_ONLY */

Since this is the effect of your "target/s390x: Use start-powered-off
CPUState property" patch, can you have a look please?

> 
> Thanks,
> Laurent
> 
>>> PSW=mask 0000000180000000 addr 00000000010004f0 cc 00
>>> R00=0000000000000000 R01=0000000000000000 R02=0000000000000000
>>> R03=0000000000000000
>>> R04=0000000000000000 R05=0000000000000000 R06=0000000000000000
>>> R07=0000000000000000
>>> R08=0000000000000000 R09=0000000000000000 R10=0000000000000000
>>> R11=0000000000000000
>>> R12=0000000000000000 R13=0000000000000000 R14=0000000000000000
>>> R15=00000040008006c0
>>>
>>> ../Makefile.target:153: recipe for target 'run-threadcount' failed
>>> make[2]: *** [run-threadcount] Error 1
>>
>> Bother.  I did see that failure on Travis, but assumed it was a false
>> positive because there were so many failures on master there.
>>
> 
>