Re: [PATCH] monitor: hmp_qemu_io: acquire aio contex, fix crash

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

22.04.2021 16:01, Max Reitz wrote:
> On 21.04.21 10:32, Vladimir Sementsov-Ogievskiy wrote:
>> Max reported the following bug:
>>
>> $ ./qemu-img create -f raw src.img 1G
>> $ ./qemu-img create -f raw dst.img 1G
>>
>> $ (echo '
>>     {"execute":"qmp_capabilities"}
>>     {"execute":"blockdev-mirror",
>>      "arguments":{"job-id":"mirror",
>>                   "device":"source",
>>                   "target":"target",
>>                   "sync":"full",
>>                   "filter-node-name":"mirror-top"}}
>> '; sleep 3; echo'
>>     {"execute":"human-monitor-command",
>>      "arguments":{"command-line":
>>                   "qemu-io mirror-top \"write 0 1G\""}}') \
>> | x86_64-softmmu/qemu-system-x86_64 \
>>     -qmp stdio \
>>     -blockdev file,node-name=source,filename=src.img \
>>     -blockdev file,node-name=target,filename=dst.img \
>>     -object iothread,id=iothr0 \
>>     -device virtio-blk,drive=source,iothread=iothr0
>>
>> crashes:
>>
>> 0  raise () at /usr/lib/libc.so.6
>> 1  abort () at /usr/lib/libc.so.6
>> 2  error_exit
>>     (err=<optimized out>,
>>     msg=msg@entry=0x55fbb1634790 <__func__.27> "qemu_mutex_unlock_impl")
>>     at ../util/qemu-thread-posix.c:37
>> 3  qemu_mutex_unlock_impl
>>     (mutex=mutex@entry=0x55fbb25ab6e0,
>>     file=file@entry=0x55fbb1636957 "../util/async.c",
>>     line=line@entry=650)
>>     at ../util/qemu-thread-posix.c:109
>> 4  aio_context_release (ctx=ctx@entry=0x55fbb25ab680) at ../util/async.c:650
>> 5  bdrv_do_drained_begin
>>     (bs=bs@entry=0x55fbb3a87000, recursive=recursive@entry=false,
>>     parent=parent@entry=0x0,
>>     ignore_bds_parents=ignore_bds_parents@entry=false,
>>     poll=poll@entry=true) at ../block/io.c:441
>> 6  bdrv_do_drained_begin
>>     (poll=true, ignore_bds_parents=false, parent=0x0, recursive=false,
>>     bs=0x55fbb3a87000) at ../block/io.c:448
>> 7  blk_drain (blk=0x55fbb26c5a00) at ../block/block-backend.c:1718
>> 8  blk_unref (blk=0x55fbb26c5a00) at ../block/block-backend.c:498
>> 9  blk_unref (blk=0x55fbb26c5a00) at ../block/block-backend.c:491
>> 10 hmp_qemu_io (mon=0x7fffaf3fc7d0, qdict=<optimized out>)
>>     at ../block/monitor/block-hmp-cmds.c:628
>>
>> man pthread_mutex_unlock
>> ...
>>      EPERM  The  mutex type is PTHREAD_MUTEX_ERRORCHECK or
>>      PTHREAD_MUTEX_RECURSIVE, or the mutex is a robust mutex, and the
>>      current thread does not own the mutex.
>>
>> So, thread doesn't own the mutex. And we have iothread here.
>>
>> Next, note that AIO_WAIT_WHILE() documents that ctx must be acquired
>> exactly once by caller. But where is it acquired in the call stack?
>> Seems nowhere.
>>
>> qemuio_command do acquire aio context.. But we need context acquired
>> around blk_unref as well. Let's do it.
>>
>> Reported-by: Max Reitz <mreitz@redhat.com>
>> Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
>> ---
>>   block/monitor/block-hmp-cmds.c | 7 +++++++
>>   1 file changed, 7 insertions(+)
>>
>> diff --git a/block/monitor/block-hmp-cmds.c b/block/monitor/block-hmp-cmds.c
>> index ebf1033f31..934100d0eb 100644
>> --- a/block/monitor/block-hmp-cmds.c
>> +++ b/block/monitor/block-hmp-cmds.c
>> @@ -559,6 +559,7 @@ void hmp_qemu_io(Monitor *mon, const QDict *qdict)
>>   {
>>       BlockBackend *blk;
>>       BlockBackend *local_blk = NULL;
>> +    AioContext *ctx;
>>       bool qdev = qdict_get_try_bool(qdict, "qdev", false);
>>       const char *device = qdict_get_str(qdict, "device");
>>       const char *command = qdict_get_str(qdict, "command");
>> @@ -615,7 +616,13 @@ void hmp_qemu_io(Monitor *mon, const QDict *qdict)
>>       qemuio_command(blk, command);
>>   fail:
>> +    ctx = blk_get_aio_context(blk);
>> +    aio_context_acquire(ctx);
>> +
>>       blk_unref(local_blk);
>> +
>> +    aio_context_release(ctx);
>> +
>>       hmp_handle_error(mon, err);
>>   }
> 
> Looks good.  Now I wonder about the rest of this function, though. qemuio_command() acquires the context on its own.  So the only thing left that looks a bit like it may want to have the context locked is blk_insert_bs().  Most of its callers seem to run in the BB’s native context, so they don’t have to acquire it; but blk_exp_add() has the context held around it, so... should this place, too?

Seems you are right. blk_insert_bs() calls bdrv_root_attach_child(), and bdrv_root_attach_child() is documented so "The caller must hold the AioContext lock @child_bs".

I'll see, what could be done here. adding one more section looks bad. Creating nested aio-context acquire when Paolo is working on removing aio context lock doesn't seem good too..

-- 
Best regards,
Vladimir