Re: [Qemu-devel] [PATCH v10 1/3] sPAPR: Implement EEH RTAS calls

[Alexander Graf <agraf@suse.de>]

On 25.06.14 01:38, Gavin Shan wrote:
> On Tue, Jun 24, 2014 at 04:43:23PM +0200, Alexander Graf wrote:
>> On 10.06.14 04:03, Gavin Shan wrote:
>>> The emulation for EEH RTAS requests from guest isn't covered
>>> by QEMU yet and the patch implements them.
>>>
>>> The patch defines constants used by EEH RTAS calls and adds
>>> callback sPAPRPHBClass::eeh_handler, which is going to be used
>>> this way:
>>>
>>> 1. RTAS calls are received in spapr_pci.c, sanity check is done
>>>     there.
>>> 2. RTAS handlers handle what they can. If there is something it
>>>     cannot handle and sPAPRPHBClass::eeh_handler callback is defined,
>>>     it is called.
>>> 3. sPAPRPHBClass::eeh_handler is only implemented for VFIO now. It
>>>     does ioctl() to the IOMMU container fd to complete the call. Error
>>>     codes from that ioctl() are transferred back to the guest.
>>>
>>> Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
>>> ---
>>>   hw/ppc/spapr_pci.c          | 248 ++++++++++++++++++++++++++++++++++++++++++++
>>>   include/hw/pci-host/spapr.h |   7 ++
>>>   include/hw/ppc/spapr.h      |  33 ++++++
>>>   3 files changed, 288 insertions(+)
>>>
>>> diff --git a/hw/ppc/spapr_pci.c b/hw/ppc/spapr_pci.c
>>> index 131434b..bfea488 100644
>>> --- a/hw/ppc/spapr_pci.c
>>> +++ b/hw/ppc/spapr_pci.c
>>> @@ -422,6 +422,241 @@ static void rtas_ibm_query_interrupt_source_number(PowerPCCPU *cpu,
>>>       rtas_st(rets, 2, 1);/* 0 == level; 1 == edge */
>>>   }
>>> +static int rtas_handle_eeh_request(sPAPRPHBState *sphb,
>>> +                                   uint32_t req, uint32_t opt,
>>> +                                   target_ulong rets)
>>> +{
>>> +    sPAPRPHBClass *info = SPAPR_PCI_HOST_BRIDGE_GET_CLASS(sphb);
>>> +    int ret;
>>> +
>>> +    ret = info->eeh_handler(sphb, req, opt);
>>> +    if (ret >= 0) {
>>> +        rtas_st(rets, 0, RTAS_OUT_SUCCESS);
>>> +    } else {
>>> +        rtas_st(rets, 0, RTAS_OUT_PARAM_ERROR);
>>> +    }
>>> +
>>> +    return ret;
>> I think you're better off failing the eeh_handler() call in here when
>> the function is not implemented. That way all callers don't have to
>> explicitly check for it.
>>
> Yep, it makes the logic of the function complete. I'll fix in v11.
>
>> I also find it a lot less confusing if the return value doesn't get
>> set inside this helper.
>>
> Yeah, I agree and will fix in v11.
>
>>> +}
>>> +
>>> +static void rtas_ibm_set_eeh_option(PowerPCCPU *cpu,
>>> +                                    sPAPREnvironment *spapr,
>>> +                                    uint32_t token, uint32_t nargs,
>>> +                                    target_ulong args, uint32_t nret,
>>> +                                    target_ulong rets)
>>> +{
>>> +    uint32_t addr, option;
>>> +    uint64_t buid = ((uint64_t)rtas_ld(args, 1) << 32) | rtas_ld(args, 2);
>>> +    sPAPRPHBState *sphb = spapr_find_phb(spapr, buid);
>>> +    sPAPRPHBClass *info = SPAPR_PCI_HOST_BRIDGE_GET_CLASS(sphb);
>>> +
>>> +    if (!sphb || !info->eeh_handler) {
>>> +        goto param_error_exit;
>>> +    }
>>> +
>>> +    if ((nargs != 4) || (nret != 1)) {
>>> +        goto param_error_exit;
>>> +    }
>>> +
>>> +    addr = rtas_ld(args, 0);
>>> +    option = rtas_ld(args, 3);
>>> +    switch (option) {
>>> +    case RTAS_EEH_ENABLE:
>>> +        if (!find_dev(spapr, buid, addr)) {
>>> +            goto param_error_exit;
>>> +        }
>>> +        break;
>>> +    case RTAS_EEH_DISABLE:
>>> +    case RTAS_EEH_THAW_IO:
>>> +    case RTAS_EEH_THAW_DMA:
>>> +        break;
>>> +    default:
>>> +        goto param_error_exit;
>>> +    }
>>> +
>>> +    rtas_handle_eeh_request(sphb, RTAS_EEH_REQ_SET_OPTION, option, rets);
>>> +    return;
>>> +
>>> +param_error_exit:
>>> +    rtas_st(rets, 0, RTAS_OUT_PARAM_ERROR);
>>> +}
>>> +
>>> +static void rtas_ibm_get_config_addr_info2(PowerPCCPU *cpu,
>>> +                                           sPAPREnvironment *spapr,
>>> +                                           uint32_t token, uint32_t nargs,
>>> +                                           target_ulong args, uint32_t nret,
>>> +                                           target_ulong rets)
>>> +{
>>> +    uint32_t addr, option;
>>> +    uint64_t buid = ((uint64_t)rtas_ld(args, 1) << 32) | rtas_ld(args, 2);
>>> +    sPAPRPHBState *sphb = spapr_find_phb(spapr, buid);
>>> +    sPAPRPHBClass *info = SPAPR_PCI_HOST_BRIDGE_GET_CLASS(sphb);
>>> +    PCIDevice *pdev;
>>> +
>>> +    if (!sphb || !info->eeh_handler) {
>>> +        goto param_error_exit;
>>> +    }
>>> +
>>> +    if ((nargs != 4) || (nret != 2)) {
>>> +        goto param_error_exit;
>>> +    }
>>> +
>>> +    addr = rtas_ld(args, 0);
>>> +    option = rtas_ld(args, 3);
>>> +    if (option != RTAS_GET_PE_ADDR && option != RTAS_GET_PE_MODE) {
>>> +        goto param_error_exit;
>>> +    }
>>> +
>>> +    pdev = find_dev(spapr, buid, addr);
>>> +    if (!pdev) {
>>> +        goto param_error_exit;
>>> +    }
>>> +
>>> +    /*
>>> +     * For now, we always have bus level PE whose address
>>> +     * has format "00BBSS00". The guest OS might regard
>>> +     * PE address 0 as invalid. We avoid that simply by
>>> +     * extending it with one.
>>> +     */
>>> +    rtas_st(rets, 0, RTAS_OUT_SUCCESS);
>>> +    if (option == RTAS_GET_PE_ADDR) {
>>> +        rtas_st(rets, 1, (pci_bus_num(pdev->bus) << 16) + 1);
>>> +    } else {
>>> +        rtas_st(rets, 1, RTAS_PE_MODE_SHARED);
>>> +    }
>>> +
>>> +    return;
>>> +
>>> +param_error_exit:
>>> +    rtas_st(rets, 0, RTAS_OUT_PARAM_ERROR);
>>> +}
>>> +
>>> +static void rtas_ibm_read_slot_reset_state2(PowerPCCPU *cpu,
>>> +                                            sPAPREnvironment *spapr,
>>> +                                            uint32_t token, uint32_t nargs,
>>> +                                            target_ulong args, uint32_t nret,
>>> +                                            target_ulong rets)
>>> +{
>>> +    uint64_t buid = ((uint64_t)rtas_ld(args, 1) << 32) | rtas_ld(args, 2);
>>> +    sPAPRPHBState *sphb = spapr_find_phb(spapr, buid);
>>> +    sPAPRPHBClass *info = SPAPR_PCI_HOST_BRIDGE_GET_CLASS(sphb);
>>> +    int32_t ret;
>>> +
>>> +    if (!sphb || !info->eeh_handler) {
>>> +        goto param_error_exit;
>>> +    }
>>> +
>>> +    if ((nargs != 3) || (nret != 4 && nret != 5)) {
>>> +        goto param_error_exit;
>>> +    }
>>> +
>>> +    ret = rtas_handle_eeh_request(sphb, RTAS_EEH_REQ_GET_STATE, 0, rets);
>>> +    if (ret >= 0) {
>>> +        rtas_st(rets, 1, ret);
>>> +        rtas_st(rets, 2, RTAS_EEH_SUPPORT);
>>> +        rtas_st(rets, 3, RTAS_EEH_PE_UNAVAIL_INFO);
>>> +        if (nret >= 5) {
>>> +            rtas_st(rets, 4, RTAS_EEH_PE_RECOVER_INFO);
>>> +        }
>>> +    }
>> It's really awkward to not see an rtas_st(0) here.
>>
> Yeah. As you mentioned above, "rtas_st(0)" would make the function
> complete (from view of logic). Thanks and I'll fix in v11.
>
>>> +
>>> +    return;
>>> +
>>> +param_error_exit:
>>> +    rtas_st(rets, 0, RTAS_OUT_PARAM_ERROR);
>>> +}
>>> +
>>> +static void rtas_ibm_set_slot_reset(PowerPCCPU *cpu,
>>> +                                    sPAPREnvironment *spapr,
>>> +                                    uint32_t token, uint32_t nargs,
>>> +                                    target_ulong args, uint32_t nret,
>>> +                                    target_ulong rets)
>>> +{
>>> +    uint32_t option;
>>> +    uint64_t buid = ((uint64_t)rtas_ld(args, 1) << 32) | rtas_ld(args, 2);
>>> +    sPAPRPHBState *sphb = spapr_find_phb(spapr, buid);
>>> +    sPAPRPHBClass *info = SPAPR_PCI_HOST_BRIDGE_GET_CLASS(sphb);
>>> +
>>> +    if (!sphb || !info->eeh_handler) {
>>> +        goto param_error_exit;
>>> +    }
>>> +
>>> +    if ((nargs != 4) || (nret != 1)) {
>>> +        goto param_error_exit;
>>> +    }
>>> +
>>> +    option = rtas_ld(args, 3);
>>> +    if (option != RTAS_SLOT_RESET_DEACTIVATE &&
>>> +        option != RTAS_SLOT_RESET_HOT &&
>>> +        option != RTAS_SLOT_RESET_FUNDAMENTAL) {
>>> +        goto param_error_exit;
>>> +    }
>> Please make this a switch().
>>
> Sure, but may I ask "switch()" could improve the code readability
> or save a bit .text space, or it's easy to be extended in future? :-)

It's mostly just readability :). It's a lot easier to understand a 
switch() at a glimpse than ANDed option != foo statements. The resulting 
assembly code should be pretty much identical.


Alex