[Sheepdog] [Qemu-devel] Re: [RFC PATCH v3 2/3] block: call the snapshot handlers of the protocol drivers

[Kevin Wolf]

Am 17.05.2010 14:19, schrieb MORITA Kazutaka:
> At Mon, 17 May 2010 13:08:08 +0200,
> Kevin Wolf wrote:
>>
>> Am 17.05.2010 12:19, schrieb MORITA Kazutaka:
>>>  
>>>  int bdrv_snapshot_goto(BlockDriverState *bs,
>>>                         const char *snapshot_id)
>>>  {
>>>      BlockDriver *drv = bs->drv;
>>> +    int ret, open_ret;
>>> +
>>>      if (!drv)
>>>          return -ENOMEDIUM;
>>> -    if (!drv->bdrv_snapshot_goto)
>>> -        return -ENOTSUP;
>>> -    return drv->bdrv_snapshot_goto(bs, snapshot_id);
>>> +    if (drv->bdrv_snapshot_goto)
>>> +        return drv->bdrv_snapshot_goto(bs, snapshot_id);
>>> +
>>> +    if (bs->file) {
>>> +        drv->bdrv_close(bs);
>>> +        ret = bdrv_snapshot_goto(bs->file, snapshot_id);
>>> +        open_ret = drv->bdrv_open(bs, bs->open_flags);
>>> +        if (open_ret < 0) {
>>> +            bdrv_delete(bs);
>>
>> I think you mean bs->file here.
>>
>> Kevin
> 
> This is an error of re-opening the format driver, so what we should
> delete here is not bs->file but bs, isn't it?  If we failed to open bs
> here, the drive doesn't seem to work anymore.

But bdrv_delete means basically free it. This is almost guaranteed to
lead to crashes because that BlockDriverState is still in use in other
places.

One additional case of use after free is in the very next line:

>>> +            bs->drv = NULL;

You can't do that when bs is freed, obviously. But I think just setting
bs->drv to NULL without bdrv_deleting it before is the better way. It
will fail any requests (with -ENOMEDIUM), but can't produce crashes.
This is also what bdrv_commit does in such situations.

In this state, we don't access the underlying file any more, so we could
delete bs->file - this is why I thought you actually meant to do that.

Kevin