Re: [Qemu-devel] [Qemu-ppc] [PATCH 4/5] target-ppc: Handle ibm, nmi-register RTAS call

[Aravinda Prasad]

On Friday 05 September 2014 02:16 PM, Alexander Graf wrote:
> 
> 
> On 04.09.14 15:49, Aravinda Prasad wrote:
>>
>>
>> On Thursday 04 September 2014 06:39 PM, Alexander Graf wrote:
>>>
>>>
>>>> Am 04.09.2014 um 10:25 schrieb Aravinda Prasad <address@hidden>:
>>>>
>>>>
>>>>
>>>>> On Friday 29 August 2014 03:46 AM, Alexander Graf wrote:
>>>>>
>>>>>
>>>>>> On 28.08.14 19:42, Aravinda Prasad wrote:
>>>>>>
>>>>>>
>>>>>>> On Thursday 28 August 2014 02:07 PM, Alexander Graf wrote:
>>>>>>>
>>>>>>>
>>>>>>>> On 28.08.14 08:38, Aravinda Prasad wrote:
>>>>>>>>
>>>>>>>>
>>>>>>>>> On Wednesday 27 August 2014 04:07 PM, Alexander Graf wrote:
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>> On 25.08.14 15:45, Aravinda Prasad wrote:
>>>>>>>>>> This patch adds FWNMI support in qemu for powerKVM
>>>>>>>>>> guests by handling the ibm,nmi-register rtas call.
>>>>>>>>>> Whenever OS issues ibm,nmi-register RTAS call, the
>>>>>>>>>> machine check notification address is saved and the
>>>>>>>>>> machine check interrupt vector 0x200 is patched to
>>>>>>>>>> issue a private hcall.
>>>>>>>>>>
>>>>>>>>>> Signed-off-by: Aravinda Prasad <address@hidden>
>>>>>>>>>> ---
>>>>>>>>>> hw/ppc/spapr_rtas.c    |   91 
>>>>>>>>>> ++++++++++++++++++++++++++++++++++++++++++++++++
>>>>>>>>>> include/hw/ppc/spapr.h |    8 ++++
>>>>>>>>>> 2 files changed, 98 insertions(+), 1 deletion(-)
>>>>>>>>>>
>>>>>>>>>> diff --git a/hw/ppc/spapr_rtas.c b/hw/ppc/spapr_rtas.c
>>>>>>>>>> index 02ddbf9..1135d2b 100644
>>>>>>>>>> --- a/hw/ppc/spapr_rtas.c
>>>>>>>>>> +++ b/hw/ppc/spapr_rtas.c
>>>>>>>>>> @@ -277,6 +277,91 @@ static void 
>>>>>>>>>> rtas_ibm_set_system_parameter(PowerPCCPU *cpu,
>>>>>>>>>>     rtas_st(rets, 0, ret);
>>>>>>>>>> }
>>>>>>>>>>
>>>>>>>>>> +static void rtas_ibm_nmi_register(PowerPCCPU *cpu,
>>>>>>>>>> +                                  sPAPREnvironment *spapr,
>>>>>>>>>> +                                  uint32_t token, uint32_t nargs,
>>>>>>>>>> +                                  target_ulong args,
>>>>>>>>>> +                                  uint32_t nret, target_ulong rets)
>>>>>>>>>> +{
>>>>>>>>>> +    int i;
>>>>>>>>>> +    uint32_t branch_inst = 0x48000002;
>>>>>>>>>> +    target_ulong guest_machine_check_addr;
>>>>>>>>>> +    PowerPCCPUClass *pcc = POWERPC_CPU_GET_CLASS(cpu);
>>>>>>>>>> +    /*
>>>>>>>>>> +     * Trampoline saves r3 in sprg2 and issues private hcall
>>>>>>>>>> +     * to request qemu to build error log. QEMU builds the
>>>>>>>>>> +     * error log, copies to rtas-blob and returns the address.
>>>>>>>>>> +     * The initial 16 bytes in rtas-blob consists of saved srr0
>>>>>>>>>> +     * and srr1 which we restore and pass on the actual error
>>>>>>>>>> +     * log address to OS handled mcachine check notification
>>>>>>>>>> +     * routine
>>>>>>>>>> +     */
>>>>>>>>>> +    uint32_t trampoline[] = {
>>>>>>>>>> +        0x7c7243a6,    /* mtspr   SPRN_SPRG2,r3 */
>>>>>>>>>> +        0x38600000,    /* li      r3,0   */
>>>>>>>>>> +        /* 0xf004 is the KVMPPC_H_REPORT_ERR private HCALL */
>>>>>>>>>> +        0x6063f004,    /* ori     r3,r3,f004  */
>>>>>>>>>> +        /* Issue H_CALL */
>>>>>>>>>> +        0x44000022,    /*  sc      1     */
>>>>>>>>>
>>>>>>>>> So up to here we're saving r3 in SPRG2 (how do we know that we can
>>>>>>>>> clobber it?) and call our special hypercall.
>>>>>>>>>
>>>>>>>>> But what does all the cruft below here do?
>>>>>>>>
>>>>>>>> The saved r3 in SPRG2 is consumed in KVMPPC_H_REPORT_ERR hcall, hence 
>>>>>>>> we
>>>>>>>> can clobber SPRG2 after hcall returns. I have included a comment in
>>>>>>>> patch 3/5 while building error log. I think better I add one here as 
>>>>>>>> well.
>>>>>>>>
>>>>>>>>>
>>>>>>>>>> +        0x7c9243a6,    /* mtspr r4 sprg2 */
>>>>>>>>>
>>>>>>>>> Apart from th fact that your order is wrong, this destroys the value 
>>>>>>>>> of
>>>>>>>>> r3 that we saved above again.
>>>>>>>>
>>>>>>>> SPRG2 is saved inside hcall and hence we don't need SPRG2 further after
>>>>>>>> KVMPPC_H_REPORT_ERR hcall returns.
>>>>>>>>
>>>>>>>>>
>>>>>>>>>> +        0xe8830000,    /* ld r4, 0(r3) */
>>>>>>>>>> +        0x7c9a03a6,    /* mtspr r4, srr0 */
>>>>>>>>>> +        0xe8830008,    /* ld r4, 8(r3) */
>>>>>>>>>> +        0x7c9b03a6,    /* mtspr r4, srr1 */
>>>>>>>>>
>>>>>>>>> Can't we just set srr0 and srr1 directly?
>>>>>>>>
>>>>>>>> I checked for instructions in ISA which set srr0/1 directly given an
>>>>>>>> address, but could not find any such instructions.
>>>>>>>
>>>>>>> I mean from QEMU :).
>>>>>>
>>>>>> srr0 and srr1, which are properly set when 0x200 is invoked, are
>>>>>> clobbered when we return from KVMPPC_H_REPORT_ERR hcall. I think they
>>>>>> are modified before issuing rfid (I can see them getting clobbered from
>>>>>> QEMU monitor). However when we jump to OS registered machine check
>>>>>> routine srr0 and srr1 should reflect the value they had when 0x200 was
>>>>>> invoked.
>>>>>>
>>>>>> Hence srr0 and srr1 are saved in hcall and restored when we return from
>>>>>> hcall. Also we don't have enough scratch registers available to save
>>>>>> these before invoking hcall from 0x200.
>>>>>>
>>>>>> Or am I missing other ways to do this from QEMU?
>>>>>
>>>>> If you just do
>>>>>
>>>>> cpu_synchronize_state() and then change env->spr[SPRN_SRR0/1] inside
>>>>> your hypercall handler that should also change the value when you return
>>>>> from the hcall.
>>>>
>>>> I tried cpu_synchronize_state(), however, srr0 and srr1 are still 
>>>> clobbered.
>>>>
>>>> Just before I issue hcall from 0x200 I see the following values from
>>>> QEMU monitor:
>>>>
>>>> SRR0 d000000000f40264  SRR1 8000000000209033
>>>>
>>>> Inside hcall, I call cpu_synchronize_state(). As soon as I return from
>>>> hcall I see:
>>>>
>>>> SRR0 0000000000000214  SRR1 8000000000001001
>>>>
>>>> I see SRR0 is now set to nip in 0x200 and SRR1 to msr value. I think it
>>>> is reset during returning from hcall before executing rfid.
> 
> Ah, because the hypercall is an interrupt itself. Heh ;). True.

Yes. However, as you mentioned, we still need cpu_synchronize_state() to
avoid stale value when we read env->spr[SPR_SPRG2]

Aravinda.

> 
> 
> Alex
> 

-- 
Regards,
Aravinda