Il 16/12/2012 18:04, Michael S. Tsirkin ha scritto:
> On Thu, Dec 13, 2012 at 09:54:23AM +0100, Paolo Bonzini wrote:
>> Il 12/12/2012 22:27, Michael S. Tsirkin ha scritto:
>>>>> Maybe it's obvious to you that qdev_reset_all(x)
>>>>> does a soft reset and what soft reset means
>>>>> for each bus type
>>>>
>>>> We can define soft reset to be *independent* of the bus type. As you
>>>> said, you access it with a device register.
>>>
>>> I think qemu has one type of reset ATM which is the hard reset.
>>
>> A hard reset would kill BARs and configuration space too.
>> qdev_reset_all doesn't. Ergo, it is not a hard reset.
>>
>> But hey, I'm not wed to names. Let's call it device-level reset and
>> bus-level reset. Whatever.
>
> It's not a question of a name.
>
> ATM qemu supports one kind of reset because it's a kind of reset all
> hardware supports. The moment we start inventing
> our own one we need to document exactly what it means.
We have two, DeviceClass's and BusClass's reset members.
I'll make a patch to document them.
> You have a point about a problem.
>
> My problem is with the solution, this solution depends on the exact
> modeling for correctness and that I don't want to do since we seem to be
> re-shuffling what inherits what so often.
>
> For example I could not figure out how the reset function for virtio pci
> (which clears pending msix vectors so is required) was called by this.
The same way a PCI bus reset clears pending MSIX vectors.
qdev_reset_all(pci_dev)
-> qdev_walk_children(pci_dev, qdev_reset_one, qbus_reset_one, NULL);
-> qdev_reset_one(pci_dev, NULL);
-> device_reset(pci_dev);
-> calls dc->reset member set for virtio-*-pci, i.e. virtio_pci_reset
> Another thing that bothers me is that during regular PCI bus reset
> virtio does not invoke qdev_reset_all but with this reset, it does.
> Inconsistent.
It does.
A PCI bus reset (or FLR) calls pci_device_reset which does this
void pci_device_reset(PCIDevice *dev)
{
int r;
qdev_reset_all(&dev->qdev);
...
}
This is exactly how a PCI bus reset clears pending MSIX vectors.
Paolo
