On Tue, Dec 16, 2014 at 01:08:52AM +0100, Alexander Graf wrote:
>
>
>On 16.12.14 00:29, Gavin Shan wrote:
>> On Tue, Dec 16, 2014 at 12:13:03AM +0100, Alexander Graf wrote:
>>> On 16.12.14 00:08, Gavin Shan wrote:
>>>> On Mon, Dec 15, 2014 at 03:52:17PM +0100, Alexander Graf wrote:
>>>>> On 15.12.14 01:15, 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:
>>>>>>
>>>>>> * RTAS calls are received in spapr_pci.c, sanity check is done
>>>>>> there.
>>>>>> * RTAS handlers handle what they can. If there is something it
>>>>>> cannot handle and sPAPRPHBClass::eeh_handler callback is defined,
>>>>>> it is called.
>>>>>> * 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.
>>>>>>
>>>>>> [aik: defined RTAS tokens for EEH RTAS calls]
>>>>>> Signed-off-by: Gavin Shan <gws...@linux.vnet.ibm.com>
>>>>>> ---
>>>>>> hw/ppc/spapr_pci.c | 246
>>>>>> ++++++++++++++++++++++++++++++++++++++++++++
>>>>>> include/hw/pci-host/spapr.h | 7 ++
>>>>>> include/hw/ppc/spapr.h | 43 +++++++-
>>>>>> 3 files changed, 294 insertions(+), 2 deletions(-)
>>>>>>
>>>>>> diff --git a/hw/ppc/spapr_pci.c b/hw/ppc/spapr_pci.c
>>>>>> index 3d70efe..3bb1971 100644
>>>>>> --- a/hw/ppc/spapr_pci.c
>>>>>> +++ b/hw/ppc/spapr_pci.c
>>>>>> @@ -406,6 +406,233 @@ 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(sPAPREnvironment *spapr,
>>>>>> + uint64_t buid, uint32_t req,
>>>>>> uint32_t opt)
>>>>>> +{
>>>>>> + sPAPRPHBState *sphb = spapr_pci_find_phb(spapr, buid);
>>>>>> + sPAPRPHBClass *info = SPAPR_PCI_HOST_BRIDGE_GET_CLASS(sphb);
>>>>>
>>>>> What happens when you try to cast NULL? Could a guest process invoke a
>>>>> host assert() through this and abort the whole VM?
>>>>>
>>>>
>>>> Yes, it would cause core dump. I had one experiment to force assigning
>>>> NULL to "sphb" before doing the cast, the whole VM is aborted. So I
>>>> guess you're happy to have something as follows. If you're not suggesting
>>>> something else, I'll update the code as follows in next version:
>>>>
>>>> sPAPRPHBState *sphb = spapr_pci_find_phb(spapr, buid);
>>>> sPAPRPHBClass *info;
>>>>
>>>> if (!sphb) {
>>>> return -ENODEV;
>>>> }
>>>>
>>>> info = SPAPR_PCI_HOST_BRIDGE_GET_CLASS(sphb);
>>>> if (!info->eeh_handler) {
>>>> return -ENOENT;
>>>> }
>>>>
>>>> return info->eeh_handler(sphb, req, opt);
>>>
>>> Yes, I think this is a lot safer. And yes, the other patch looks sane to me.
>>>
>>
>> Thank you for your time reviewing this. Will update in next version.
>>
>>>>
>>>>>> +
>>>>>> + if (!sphb || !info->eeh_handler) {
>>>>>> + return -ENOENT;
>>>>>> + }
>>>>>> +
>>>>>> + return info->eeh_handler(sphb, req, opt);
>>>>>> +}
>>>>>> +
>>>>>> +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);
>>>>>> + int ret;
>>>>>> +
>>>>>> + 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 (!spapr_pci_find_dev(spapr, buid, addr)) {
>>>>>> + goto param_error_exit;
>>>>>> + }
>>>>>> + break;
>>>>>> + case RTAS_EEH_DISABLE:
>>>>>> + case RTAS_EEH_THAW_IO:
>>>>>> + case RTAS_EEH_THAW_DMA:
>>>>>
>>>>> So these don't use the addr hint?
>>>>>
>>>>
>>>> No, there're no address as argument of this RTAS call "ibm,set-eeh-option".
>>>> The RTAS call has 4 arguments, all of them are 32-bits: BUID high part,
>>>> BUID
>>>> low part, PE address, option. The option could be one of: enable/disable
>>>> EEH
>>>> functionality, enable IO path, enable DMA path.
>>>
>>> Well, I'm just wondering that ENABLE wants to make sure there's a device
>>> and the others don't.
>>>
>>
>> Oh, I misunderstood your question. Yes, you're correct. All options
>> except ENABLE will have address check in rtas_handle_eeh_request()
>> where we check on "BUID" since each PHB and PE have one-to-one
>> relationship.
>>
>> ENABLE and other options are using different address: enable
>> uses config address of one specific device, but other options use PE
>> address. From guest's sides, it enables EEH capability on basis
>> of PCI device and left options are supported on basis of PE. Each
>> PE could contain one or multiple PCI devices.
>>
>> DISABLE option isn't used until now.
>
>So would DISABLE also take effect on devfn basis or would it ignore the
>first parameter?
>
>If it behaves the same as ENABLE (which would be logical), please move
>it into the same case group.
>
DISABLE takes effect on PE basis. We don't have devfn for the case and
code needn't changes. Here's the concise procedures if DISABLE needs to be
involved:
- ENABLE on basis of PCI device
- RTAS call "ibm,get-config-addr-info2" to get the PE address
- DISABLE on basis of PE, with PE address as argument.
If you don't have anything else, I would like to change the code and
send new version to you. Thanks again for your time.
Thanks,
Gavin
