Re: [Qemu-devel] [PATCH] `qdev_free` when unplug a pci device

[Wen Congyang]

At 03/10/2011 12:31 PM, Ryan Harper Write:
> * Wen Congyang <address@hidden> [2011-03-09 01:21]:
>> At 03/09/2011 02:12 PM, Ryan Harper Write:
>>> * Wen Congyang <address@hidden> [2011-03-08 23:09]:
>>>> At 03/09/2011 12:08 PM, Ryan Harper Write:
>>>>> * Wen Congyang <address@hidden> [2011-02-27 20:56]:
>>>>>> Hi Markus Armbruster
>>>>>>
>>>>>> At 02/23/2011 04:30 PM, Markus Armbruster Write:
>>>>>>> Isaku Yamahata <address@hidden> writes:
>>>>>>>
>>>>>>
>>>>>> <snip>
>>>>>>
>>>>>>>
>>>>>>> I don't think this patch is correct.  Let me explain.
>>>>>>>
>>>>>>> Device hot unplug is *not* guaranteed to succeed.
>>>>>>>
>>>>>>> For some buses, such as USB, it always succeeds immediately, i.e. when
>>>>>>> the device_del monitor command finishes, the device is gone.  Live is
>>>>>>> good.
>>>>>>>
>>>>>>> But for PCI, device_del merely initiates the ACPI unplug rain dance.  It
>>>>>>> doesn't wait for the dance to complete.  Why?  The dance can take an
>>>>>>> unpredictable amount of time, including forever.
>>>>>>>
>>>>>>> Problem: Subsequent device_add can fail if it reuses the qdev ID or PCI
>>>>>>> slot, and the unplug has not yet completed (race condition), or it
>>>>>>> failed.  Yes, Virginia, PCI hotplug *can* fail.
>>>>>>>
>>>>>>> When unplug succeeds, the qdev is automatically destroyed.
>>>>>>> pciej_write() does that for PIIX4.  Looks like pcie_cap_slot_event()
>>>>>>> does it for PCIE.
>>>>>>
>>>>>> I got a similar problem.  When I unplug a pci device by hand, it works
>>>>>> as expected, and I can hotplug it again. But when I use a srcipt to
>>>>>> do the same thing, sometimes it failed. I think I may find another bug.
>>>>>>
>>>>>> Steps to reproduce this bug:
>>>>>> 1. cat ./test-e1000.sh # RHEL6RC is domain name
>>>>>>    #! /bin/bash
>>>>>>
>>>>>>    while true; do
>>>>>>            virsh attach-interface RHEL6RC network default --mac 
>>>>>> 52:54:00:1f:db:c7 --model e1000
>>>>>>            if [[ $? -ne 0 ]]; then
>>>>>>                    break
>>>>>>            fi
>>>>>>            virsh detach-interface RHEL6RC network --mac 52:54:00:1f:db:c7
>>>>>>            if [[ $? -ne 0 ]]; then
>>>>>>                    break
>>>>>>            fi
>>>>>>            sleep 5
>>>>>
>>>>> How do you know that the guest has responded at this point before you
>>>>> attempt to attach again at the top of the loop.  Any attach/detach
>>>>> requires the guest to respond to the request and it may not respond at
>>>>> all.
>>>>
>>>> When I attach/detach interface by hand, it works fine: I can see the new 
>>>> interface
>>>> when I attach it, and it disapears when I detached it.
>>>
>>> The point is that since the attach and detach require guest
>>> participation, this interface isn't reliable.  You have a sleep 5 in
>>> your loop, hoping to wait long enough for the guest to respond, but
>>> after a number of iterations in your loop it fails, you can bump the
>>> sleep to to 3600 seconds and the guest *still* might not respond...
>>
>> We use sci interrupt to tell the guest that a device has been 
>> attached/detached.
>> But the sci interrupt is *lost* in qemu, so the guest does not know a device 
>> has
>> been attached/detached, and does not respond it.
>>
>> If the sci interrupt is not lost, the guest can respond it.
> 
> *can* is the important word.  Even if the interrupt isn;t lost, you have
> no way to guarantee that the guest will respond at all.  That's not to
> say there isn't a bug around the lost interrupt; but rather a more
> general point about hotplug's current architecture.

I don't know whether a real hardware has the same behavior.
Should we make sure the sci interrupt not lost?

> 
>>
>>>
>>>
>