Re: Race condition in overlayed qcow2?

[Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>]

20.02.2020 13:00, Pavel Dovgalyuk wrote:
>> From: Vladimir Sementsov-Ogievskiy [mailto:vsementsov@virtuozzo.com]
>> 20.02.2020 11:31, dovgaluk wrote:
>>> Vladimir Sementsov-Ogievskiy писал 2020-02-19 19:07:
>>>> 19.02.2020 17:32, dovgaluk wrote:
>>>>> I encountered a problem with record/replay of QEMU execution and figured out the
>> following, when
>>>>> QEMU is started with one virtual disk connected to the qcow2 image with applied 'snapshot'
>> option.
>>>>>
>>>>> The patch d710cf575ad5fb3ab329204620de45bfe50caa53 "block/qcow2: introduce parallel
>> subrequest handling in read and write"
>>>>> introduces some kind of race condition, which causes difference in the data read from the
>> disk.
>>>>>
>>>>> I detected this by adding the following code, which logs IO operation checksum. And this
>> checksum may be different in different runs of the same recorded execution.
>>>>>
>>>>> logging in blk_aio_complete function:
>>>>>           qemu_log("%"PRId64": blk_aio_complete\n", replay_get_current_icount());
>>>>>           QEMUIOVector *qiov = acb->rwco.iobuf;
>>>>>           if (qiov && qiov->iov) {
>>>>>               size_t i, j;
>>>>>               uint64_t sum = 0;
>>>>>               int count = 0;
>>>>>               for (i = 0 ; i < qiov->niov ; ++i) {
>>>>>                   for (j = 0 ; j < qiov->iov[i].iov_len ; ++j) {
>>>>>                       sum += ((uint8_t*)qiov->iov[i].iov_base)[j];
>>>>>                       ++count;
>>>>>                   }
>>>>>               }
>>>>>               qemu_log("--- iobuf offset %"PRIx64" len %x sum: %"PRIx64"\n", acb-
>>> rwco.offset, count, sum);
>>>>>           }
>>>>>
>>>>> I tried to get rid of aio task by patching qcow2_co_preadv_part:
>>>>> ret = qcow2_co_preadv_task(bs, ret, cluster_offset, offset, cur_bytes, qiov, qiov_offset);
>>>>>
>>>>> That change fixed a bug, but I have no idea what to debug next to figure out the exact
>> reason of the failure.
>>>>>
>>>>> Do you have any ideas or hints?
>>>>>
>>>>
>>>> Hi!
>>>>
>>>> Hmm, do mean that read from the disk may return wrong data? It would
>>>> be very bad of course :(
>>>> Could you provide a reproducer, so that I can look at it and debug?
>>>
>>> It is just a winxp-32 image. I record the execution and replay it with the following command
>> lines:
>>>
>>> qemu-system-i386 -icount shift=7,rr=record,rrfile=replay.bin -m 512M -drive
>> file=xp.qcow2,if=none,id=device-34-file,snapshot -drive driver=blkreplay,if=none,image=device-
>> 34-file,id=device-34-driver -device ide-hd,drive=device-34-driver,bus=ide.0,id=device-34 -net
>> none
>>>
>>> qemu-system-i386 -icount shift=7,rr=replay,rrfile=replay.bin -m 512M -drive
>> file=xp.qcow2,if=none,id=device-34-file,snapshot -drive driver=blkreplay,if=none,image=device-
>> 34-file,id=device-34-driver -device ide-hd,drive=device-34-driver,bus=ide.0,id=device-34 -net
>> none
>>>
>>> Replay stalls at some moment due to the non-determinism of the execution (probably caused by
>> the wrong data read).
>>
>> Hmm.. I tried it  (with x86_64 qemu and centos image). I waited for some time for a first
>> command, than Ctrl+C it. After it replay.bin was 4M. Than started the second command. It
>> works, not failing, not finishing. Is it bad? What is expected behavior and what is wrong?
> 
> The second command should finish. There is no replay introspection yet (in master), but you can
> stop qemu with gdb and inspect replay_state.current_icount field. It should increase with every
> virtual CPU instruction execution. If that counter has stopped, it means that replay hangs.

It hangs for me even with QCOW2_MAX_WORKERS = 1..

> 
>>>> What is exactly the case? May be you have other parallel aio
>>>> operations to the same region?
>>>
>>> As far as I understand, all aio operations, initiated by IDE controller, are performed one-
>> by-one.
>>> I don't see anything else in the logs.
>>>
>>>> Ideas to experiment:
>>>>
>>>> 1. change QCOW2_MAX_WORKERS to 1 or to 2, will it help?
>>>
>>> 1 or 2 are ok, and 4 or 8 lead to the failures.
>>>
>>>> 2. understand what is the case in code: is it read from one or several
>>>> clusters, is it aligned,
>>>> what is the type of clusters, is encryption in use, compression?
>>>
>>> There is no encryption and I thinks compression is not enabled too.
>>> Clusters are read from the temporary overlay:
>>>
>>> blk_aio_prwv
>>> blk_aio_read_entry
>>> bdrv_co_preadv_part complete offset: 26300000 qiov_offset: 1c200 len: 1e00
>>> bdrv_co_preadv_part complete offset: 24723e00 qiov_offset: 0 len: 1c200
>>> bdrv_co_preadv_part complete offset: c0393e00 qiov_offset: 0 len: 1e000
>>> bdrv_co_preadv_part complete offset: c0393e00 qiov_offset: 0 len: 1e000
>>> bdrv_co_preadv_part complete offset: c0393e00 qiov_offset: 0 len: 1e000
>>>
>>>
>>>> 3. understand what kind of data corruption. What we read instead of
>>>> correct data? Just garbage, or may be zeroes, or what..
>>>
>>> Most bytes are the same, but some are different:
>>>
>>> < 00 00 00 00 00 00 00 00 00 00 00 00 00 00 02 00
>>> < 46 49 4c 45 30 00 03 00 18 d1 33 02 00 00 00 00
>>> < 01 00 01 00 38 00 01 00 68 01 00 00 00 04 00 00
>>> < 00 00 00 00 00 00 00 00 04 00 00 00 9d 0e 00 00
>>> < 02 00 00 00 00 00 00 00 10 00 00 00 60 00 00 00
>>> ---
>>>> 00 00 00 00 00 00 00 00 00 00 00 00 00 00 04 00
>>>> 46 49 4c 45 30 00 03 00 86 78 35 03 00 00 00 00
>>>> 01 00 01 00 38 00 01 00 60 01 00 00 00 04 00 00
>>>> 00 00 00 00 00 00 00 00 04 00 00 00 a1 0e 00 00
>>>> 04 00 00 00 00 00 00 00 10 00 00 00 60 00 00 00
>>>
>>> That is strange. I could think, that it was caused by the bugs in
>>> deterministic CPU execution, but the first difference in logs
>>> occur in READ operation (I dump read/write buffers in blk_aio_complete).
>>>
>>
>> Aha, yes, looks strange.
>>
>> Then next steps:
>>
>> 1. Does problem hit into the same offset every time?
> 
> Yes, almost the same offset, almost the same phase of the execution.
> 
>> 2. Do we write to this region before this strange read?
> 
> No.
> 
>> 2.1. If yes, we need to check that we read what we write.. You say you dump buffers
>> in blk_aio_complete... I think it would be more reliable to dump at start of
>> bdrv_co_pwritev and at end of bdrv_co_preadv. Also, guest may modify its buffers
>> during operation which would be strange but possible.
> 
> I dumped every write in file-posix.c handle_aiocb_rw_linear and qemu_pwritev
> and found no difference in executions.
> 
>> 2.2 If not, hmm...
> 
> Exactly.
> 
> Pavel Dovgalyuk
> 


-- 
Best regards,
Vladimir