Re: [Qemu-devel] How to emulate block I/O timeout on qemu side?

[John Snow]

On 11/03/2018 01:24 PM, Dongli Zhang wrote:
> Hi all,
> 

Hi, please reply below the quoted text when writing to qemu-devel in the
future; my reply is below.

> I tried with the patch at:
> 
> https://lists.gnu.org/archive/html/qemu-devel/2018-09/msg00394.html
> 
> The patch is applied to qemu-3.0.0.
> 
> 
> Below configuration is used to test the feature for guest VM nvme.
> 
> # qemu-system-x86_64 \
> -smp 4 -m 2000M -enable-kvm -vnc :0 -monitor stdio \
> -net nic -net user,hostfwd=tcp::5022-:22 \
> -drive file=virtio-disk.img,format=raw,if=none,id=disk0 \
> -device virtio-blk-pci,drive=disk0,id=disk0-dev,num-queues=2,iothread=io1 \
> -object iothread,id=io1 \
> -device nvme,drive=nvme1,serial=deadbeaf1 \
> -drive file=blkdebug:blkdebug.config:nvme.img,if=none,id=nvme1
> 
> # cat blkdebug.config
> [delay]
> event = "write_aio"
> latency = "9999999999"
> sector = "40960"
> 
> 
> The 'write' latency of sector=40960 is set to a very large value. When the I/O
> is stalled in guest due to that sector=40960 is accessed, I do see below
> messages in guest log:
> 
> [   80.807755] nvme nvme0: I/O 11 QID 2 timeout, aborting
> [   80.808095] nvme nvme0: Abort status: 0x4001
> 
> 
> However, then nothing happens further. nvme I/O hangs in guest. I am not able 
> to
> kill the qemu process with Ctrl+C. Both vnc and qemu user net do not work. I
> need to kill qemu with "kill -9"
> >
> The same result for virtio-scsi and qemu is stuck as well.
> 

OK, sounds like a bug in the delay implementation here, then; or
something I've not considered with the locking/drain specifics. Thanks
for the report.

> 
> About blkdebug, I can only trigger the error by the config file. Is there a 
> way
> to inject error or latency via qemu monior? For instance, I would like to 
> inject
> error not for a specific sector or state, but for the entire disk when I input
> some command via qemu monitor.
> 

I don't recall.

There are some tricks you can play with set-state and rules that only
apply when in a certain state. I don't remember if there are monitor or
QMP commands to set the state explicitly.

I'm looking at docs/devel/blkdebug.txt and don't see anything immediately.

There's maybe a way you can use blockdev-add to create the blkdebug node
and insert it live into the graph when you want it, and live-remove it
when you don't, but I'm not sure of the syntax right away.

(maybe that's not possible?)

--js

> Dongli Zhang
> 
> 
> On 11/03/2018 02:17 AM, John Snow wrote:
>>
>>
>> On 11/02/2018 01:55 PM, Marc Olson wrote:
>>> On 11/2/18 10:49 AM, John Snow wrote:
>>>> On 11/02/2018 04:11 AM, Dongli Zhang wrote:
>>>>> Hi,
>>>>>
>>>>> Is there any way to emulate I/O timeout on qemu side (not fault
>>>>> injection in VM
>>>>> kernel) without modifying qemu source code?
>>>>>
>>>>> For instance, I would like to observe/study/debug the I/O timeout
>>>>> handling of
>>>>> nvme, scsi, virtio-blk (not supported) of VM kernel.
>>>>>
>>>>> Is there a way to trigger this on purpose on qemu side?
>>>>>
>>>>> Thank you very much!
>>>>>
>>>>> Dongli Zhang
>>>>>
>>>> I don't think the blkdebug driver supports arbitrary delays right now.
>>>> Maybe we could augment it to do so?
>>>>
>>>> (I thought someone already had, but maybe it wasn't merged?)
>>>>
>>>> Aha, here:
>>>>
>>>> https://lists.gnu.org/archive/html/qemu-devel/2018-08/msg05297.html
>>>> V2: https://lists.gnu.org/archive/html/qemu-devel/2018-09/msg00394.html
>>>>
>>>> Let's work from there.
>>>
>>> I've got updates to that patch series that fell on the floor due to
>>> other competing things. I'll get some screen time this weekend to work
>>> on them and submit v3.
>>>
>>> /marc
>>>
>>
>> Great! Please CC the usual maintainers, but also include me.
>>
>> In the meantime, Dongli Zhang, why don't you try the v2 patch and see if
>> that helps you out for your use case? Report back if it works for you or
>> not.
>>
>> --js
>>