Re: [BUG, RFC] cpr-transfer: qxl guest driver crashes after migration

[Steven Sistare <steven.sistare@oracle.com>]

On 2/28/2025 12:39 PM, Andrey Drobyshev wrote:
> Hi all,
> 
> We've been experimenting with cpr-transfer migration mode recently and
> have discovered the following issue with the guest QXL driver:
> 
> Run migration source:
>> EMULATOR=/path/to/emulator
>> ROOTFS=/path/to/image
>> QMPSOCK=/var/run/alma8qmp-src.sock
>>
>> $EMULATOR -enable-kvm \
>>      -machine q35 \
>>      -cpu host -smp 2 -m 2G \
>>      -object memory-backend-file,id=ram0,size=2G,mem-path=/dev/shm/ram0,share=on\
>>      -machine memory-backend=ram0 \
>>      -machine aux-ram-share=on \
>>      -drive file=$ROOTFS,media=disk,if=virtio \
>>      -qmp unix:$QMPSOCK,server=on,wait=off \
>>      -nographic \
>>      -device qxl-vga
> 
> Run migration target:
>> EMULATOR=/path/to/emulator
>> ROOTFS=/path/to/image
>> QMPSOCK=/var/run/alma8qmp-dst.sock
>>                                                                                     
>> $EMULATOR -enable-kvm \
>>      -machine q35 \
>>      -cpu host -smp 2 -m 2G \
>>      -object memory-backend-file,id=ram0,size=2G,mem-path=/dev/shm/ram0,share=on\
>>      -machine memory-backend=ram0 \
>>      -machine aux-ram-share=on \
>>      -drive file=$ROOTFS,media=disk,if=virtio \
>>      -qmp unix:$QMPSOCK,server=on,wait=off \
>>      -nographic \
>>      -device qxl-vga \
>>      -incoming tcp:0:44444 \
>>      -incoming '{"channel-type": "cpr", "addr": { "transport": "socket", "type": "unix", "path": "/var/run/alma8cpr-dst.sock"}}'
> 
> 
> Launch the migration:
>> QMPSHELL=/root/src/qemu/master/scripts/qmp/qmp-shell
>> QMPSOCK=/var/run/alma8qmp-src.sock
>>
>> $QMPSHELL -p $QMPSOCK <<EOF
>>      migrate-set-parameters mode=cpr-transfer
>>      migrate channels=[{"channel-type":"main","addr":{"transport":"socket","type":"inet","host":"0","port":"44444"}},{"channel-type":"cpr","addr":{"transport":"socket","type":"unix","path":"/var/run/alma8cpr-dst.sock"}}]
>> EOF
> 
> Then, after a while, QXL guest driver on target crashes spewing the
> following messages:
>> [   73.962002] [TTM] Buffer eviction failed
>> [   73.962072] qxl 0000:00:02.0: object_init failed for (3149824, 0x00000001)
>> [   73.962081] [drm:qxl_alloc_bo_reserved [qxl]] *ERROR* failed to allocate VRAM BO
> 
> That seems to be a known kernel QXL driver bug:
> 
> https://lore.kernel.org/all/20220907094423.93581-1-min_halo@163.com/T/
> https://lore.kernel.org/lkml/ZTgydqRlK6WX_b29@eldamar.lan/
> 
> (the latter discussion contains that reproduce script which speeds up
> the crash in the guest):
>> #!/bin/bash
>>
>> chvt 3
>>
>> for j in $(seq 80); do
>>          echo "$(date) starting round $j"
>>          if [ "$(journalctl --boot | grep "failed to allocate VRAM BO")" != "" ]; then
>>                  echo "bug was reproduced after $j tries"
>>                  exit 1
>>          fi
>>          for i in $(seq 100); do
>>                  dmesg > /dev/tty3
>>          done
>> done
>>
>> echo "bug could not be reproduced"
>> exit 0
> 
> The bug itself seems to remain unfixed, as I was able to reproduce that
> with Fedora 41 guest, as well as AlmaLinux 8 guest. However our
> cpr-transfer code also seems to be buggy as it triggers the crash -
> without the cpr-transfer migration the above reproduce doesn't lead to
> crash on the source VM.
> 
> I suspect that, as cpr-transfer doesn't migrate the guest memory, but
> rather passes it through the memory backend object, our code might
> somehow corrupt the VRAM.  However, I wasn't able to trace the
> corruption so far.
> 
> Could somebody help the investigation and take a look into this?  Any
> suggestions would be appreciated.  Thanks!

Possibly some memory region created by qxl is not being preserved.
Try adding these traces to see what is preserved:

-trace enable='*cpr*'
-trace enable='*ram_alloc*'

- Steve