Re: [Qemu-devel] [PATCH v2] pci: Introduce helper to retrieve a PCI device's DMA address space

[Paolo Bonzini]

Il 09/08/2013 12:58, Alexey Kardashevskiy ha scritto:
> On 08/09/2013 08:19 PM, Paolo Bonzini wrote:
>> Il 09/08/2013 12:13, Alexey Kardashevskiy ha scritto:
>>> On 08/09/2013 07:53 PM, Paolo Bonzini wrote:
>>>> Il 09/08/2013 11:48, Alexey Kardashevskiy ha scritto:
>>>>> On 08/09/2013 07:40 PM, Paolo Bonzini wrote:
>>>>>> Il 09/08/2013 10:49, Alexey Kardashevskiy ha scritto:
>>>>>>> A PCI device's DMA address space (possibly an IOMMU) is returned by a
>>>>>>> method on the PCIBus.  At the moment that only has one caller, so the
>>>>>>> method is simply open coded.  We'll need another caller for VFIO, so
>>>>>>> this patch introduces a helper/wrapper function.
>>>>>>>
>>>>>>> Signed-off-by: David Gibson <address@hidden>
>>>>>>> [aik: added inheritance from parent if iommu is not set for the current 
>>>>>>> bus]
>>>>>>> Signed-off-by: Alexey Kardashevskiy <address@hidden>
>>>>>>>
>>>>>>> ---
>>>>>>> Changes:
>>>>>>> v2:
>>>>>>> * added inheritance, needed for a pci-bridge on spapr-ppc64
>>>>>>> * pci_iommu_as renamed to pci_device_iommu_address_space
>>>>>>> ---
>>>>>>>  hw/pci/pci.c         | 22 ++++++++++++++++------
>>>>>>>  include/hw/pci/pci.h |  1 +
>>>>>>>  2 files changed, 17 insertions(+), 6 deletions(-)
>>>>>>>
>>>>>>> diff --git a/hw/pci/pci.c b/hw/pci/pci.c
>>>>>>> index 4c004f5..0072b54 100644
>>>>>>> --- a/hw/pci/pci.c
>>>>>>> +++ b/hw/pci/pci.c
>>>>>>> @@ -812,12 +812,7 @@ static PCIDevice *do_pci_register_device(PCIDevice 
>>>>>>> *pci_dev, PCIBus *bus,
>>>>>>>      }
>>>>>>>  
>>>>>>>      pci_dev->bus = bus;
>>>>>>> -    if (bus->iommu_fn) {
>>>>>>> -        dma_as = bus->iommu_fn(bus, bus->iommu_opaque, devfn);
>>>>>>> -    } else {
>>>>>>> -        /* FIXME: inherit memory region from bus creator */
>>>>>>> -        dma_as = &address_space_memory;
>>>>>>> -    }
>>>>>>> +    dma_as = pci_device_iommu_address_space(pci_dev);
>>>>>>>  
>>>>>>>      memory_region_init_alias(&pci_dev->bus_master_enable_region,
>>>>>>>                               OBJECT(pci_dev), "bus master",
>>>>>>> @@ -2239,6 +2234,21 @@ static void pci_device_class_init(ObjectClass 
>>>>>>> *klass, void *data)
>>>>>>>      k->props = pci_props;
>>>>>>>  }
>>>>>>>  
>>>>>>> +AddressSpace *pci_device_iommu_address_space(PCIDevice *dev)
>>>>>>> +{
>>>>>>> +    PCIBus *bus = PCI_BUS(dev->bus);
>>>>>>> +
>>>>>>> +    if (bus->iommu_fn) {
>>>>>>> +        return bus->iommu_fn(bus, bus->iommu_opaque, dev->devfn);
>>>>>>> +    }
>>>>>>> +
>>>>>>> +    if (bus->parent_dev) {
>>>>>>> +        return pci_device_iommu_address_space(bus->parent_dev);
>>>>>>> +    }
>>>>>>
>>>>>> No, this would fail if bus->parent_dev is not NULL but not a PCI device
>>>>>> either.
>>>>>
>>>>> parent_dev is of the PCIDevice* type, how can it be not a PCI device? :-/
>>>>
>>>> Doh, I misread the code, I thought it was the "parent" field in
>>>> BusState.  Why do we have parent_dev at all?
>>>
>>> The code is too old? Don't know.
>>>
>>>
>>>>>> You can use object_dynamic_cast to convert the parent_dev to
>>>>>> PCIDevice, and if the cast succeeds you call the new function.
>>>>>>
>>>>>> Perhaps you could make the new function take a PCIBus instead.
>>>>>> Accessing the PCIDevice's IOMMU address space (as opposed to the
>>>>>> bus-master address space) doesn't make much sense, VFIO is really a
>>>>>> special case.  Putting the new API on the bus side instead looks better.
>>>>>>
>>>>>> (BTW, do you need to enable bus-master DMA on PCI bridges, to do DMA for
>>>>>> devices sitting on the secondary bus?)
>>>>>
>>>>> It happens naturally I guess when linux enables devices.
>>>>
>>>> Yes, but then using the IOMMU address space would be wrong; you would
>>>> have to use the bus-master address space as a base for the child's
>>>> bus-master address space.
>>>
>>>
>>> Like this? Works too.
>>>
>>> diff --git a/hw/pci/pci.c b/hw/pci/pci.c
>>> index 8bdcedc..a4c70e6 100644
>>> --- a/hw/pci/pci.c
>>> +++ b/hw/pci/pci.c
>>> @@ -2247,23 +2247,23 @@ AddressSpace
>>> *pci_device_iommu_address_space(PCIDevice *dev)
>>>  {
>>>      PCIBus *bus = PCI_BUS(dev->bus);
>>>
>>>      if (bus->iommu_fn) {
>>>          return bus->iommu_fn(bus, bus->iommu_opaque, dev->devfn);
>>>      }
>>>
>>>      if (bus->parent_dev) {
>>>          return pci_device_iommu_address_space(bus->parent_dev);
>>>      }
>>>
>>> -    return &address_space_memory;
>>> +    return &dev->bus_master_as;
>>>  }
>>
>> I was thinking more like this:
>>
>>      if (bus->parent_dev) {
>> -        return pci_device_iommu_address_space(bus->parent_dev);
>> +        /* Take parent device's bus-master enable bit into account.  */
>> +        return pci_get_address_space(bus->parent_dev);
>>      }
>>
>> +    /* Not a secondary bus and no IOMMU.  Use system memory.  */
>>      return &address_space_memory;
> 
> Ok.
> 
> Oh. BTW. This "bus master" thing now breaks VFIO's check for all devices
> being in the same address space as every single device has its own "bus
> master" AddressSpace.

Yes, fixing that check is another good use of the new API you introduced.

But after Ben's answer, I guess the above change is not really needed.
It would add complication for VFIO, too.  Proper emulation would require
QEMU to trap writes to the device's bus-master bit.  QEMU would have to
take of the value that the guest writes, AND it with the bus-master
bit(s) of all bridges between the host bridge and the VFIO device, and
write the result to the passed-through device.  This is because the
bridges are emulated, and do not exist in real hardware.

>>>> Also, we would have to fix the x86 firmware.
> 
> btw what would it mean? :)

Firmware would have to look for bridges and enable bus master DMA on
them.  Otherwise, for example a SCSI controller below a bridge (not
virtio-scsi) would fail to boot.

Paolo