Re: xl: pci completion error

[Sergey Tovpeko <tsv.devel@gmail.com>]

Stefano Stabellini wrote:
> On Tue, 5 Oct 2010, Sergey Tovpeko wrote:
>   
>> Hello, list!
>>
>> xl reports the error on passthrough-ed pci device removing.
>>
>> do_pci_remove device 01:00.0
>> libxl: error: libxl_device.c:448:libxl__wait_for_device_model Device 
>> Model not ready
>> libxl: error: libxl_pci.c:858:do_pci_remove Device Model didn't respond 
>> in time
>> libxl: error: libxl.c:944:libxl_domain_destroy pci shutdown failed for 
>> domid 1
>> libxl: error: libxl.c:896:libxl_destroy_device_model Couldn't find 
>> device model's pid: No such file or directory
>> libxl: error: libxl.c:956:libxl_domain_destroy 
>> libxl_destroy_device_model failed for 1
>> libxl: error: libxl_device.c:307:libxl__devices_destroy 
>> /local/domain/1/device is empty
>>
>>
>>
>> It seems that libxl_pci didn't get the 'pci-removed' status from 
>> qemu-dm. Please, have a look  who should set this status in qemu-dm. As 
>> for me I added xenstore_record_dm_state("pci-removed");
>> after do_pci_del(par);
>> in xenstore_process_dm_command_event function.
>>
>> It fixed up my issue of removing pci devices.
>>     
>
> What guest OS are you using?
> Currently "pci-removed" is only written in response of an eject command
> from the guest OS, that means that if the guest doesn't support pci
> hotplug the value won't be written.
> If you are using Linux you should make sure that the acpiphp module is
> loaded, if you are using Windows, I think the only version that supports
> pci hotplug is Windows Server 2008 but I might be wrong.
>   

The guest is Windows XP sp2 32 bit.

When I do pci hot-unplug on the fly
xl pci-detach 11:0b.0
the guest detaches the pci card correctly with 'pci-removed' signal, but 
on the shutdown, it doesn't.

In the code the flow is the following:
qemu-dm receives the 'pci-rem' command and sends the SCI interrupt to 
the guest. I think guest should catch this interrupt and write to the 
magic ACPI port that causes qemu-dm send 'pci-removed' signal.

On guest shutdown,  libxl sends 'pci-rem' command to qemu-dm and waits 
for the answer.
I suppose, by the time 'pci-rem' signal my guest have finished its work, 
and doesn't react on SCI interrupt. As a result libxl doesn't receive 
pci-removed signal.

It's my observation.


Sergey.