Re: [Qemu-devel] Safely reopening image files by stashing fds

[Kevin Wolf]

Am 09.08.2011 11:22, schrieb supriya kannery:
> Kevin Wolf wrote:
>> Am 08.08.2011 09:02, schrieb Supriya Kannery:
>>   
>>> On 08/05/2011 09:19 PM, Anthony Liguori wrote:
>>>     
>>>> On 08/05/2011 10:43 AM, Kevin Wolf wrote:
>>>>       
>>>>> Am 05.08.2011 17:24, schrieb Stefan Hajnoczi:
>>>>>         
>>>>>> On Fri, Aug 5, 2011 at 3:28 PM, Christoph Hellwig<address@hidden> wrote:
>>>>>>           
>>>>>>> On Fri, Aug 05, 2011 at 02:12:48PM +0100, Daniel P. Berrange wrote:
>>>>>>>             
>>>>>>>>> Because you cannot change O_DIRECT on an open fd :(. This is why
>>>>>>>>> we're going through this pain.
>>>>>>>>>                 
>>>>>>>> Hmm, I remember hearing that before, but looking at the current
>>>>>>>> fcntl()
>>>>>>>> manpage, it claims you *can* change O_DIRECT using SET_FL. Perhaps
>>>>>>>> this
>>>>>>>> is a newish feature, but it'd be nicer to use it if possible ?
>>>>>>>>               
>>>>>>> It's been there since day 1 of O_DIRECT support.
>>>>>>>             
>>>>>> Sorry, my bad. So for Linux we could just use fcntl for
>>>>>> block_set_hostcache and not bother with reopening. However, we will
>>>>>> need to reopen should we wish to support changing O_DSYNC.
>>>>>>           
>>>>> We do wish to support that.
>>>>>
>>>>> Anthony thinks that allowing the guest to toggle WCE is a prerequisite
>>>>> for making cache=writeback the default. And this is something that I
>>>>> definitely want to do for 1.0.
>>>>>         
>>>> Indeed.
>>>>
>>>>       
>>> We discussed the following so far...
>>> 1. How to safely reopen image files
>>> 2. Dynamic hostcache change
>>> 3. Support for dynamic change of O_DSYNC
>>>
>>> Since 2 is independent of 1, shall I go ahead implementing
>>> hostcache change using fcntl.
>>>
>>> Implementation for safely reopening image files using "BDRVReopenState"
>>> can be done separately as a pre-requisite before implementing 3
>>>     
>>
>> Doing it separately means that we would introduce yet another callback
>> that is used just to change O_DIRECT. In the end we want it to use
>> bdrv_reopen(), too, so I'm not sure if there is a need for a temporary
>> solution.
>>
>>   
> Could you please explain "In the end we want it to use bdrv_reopen" at 
> bit more.
> When fcntl() can change O_DIRECT on  open fd , is there a  need to 
> "re-open"
> the image file?
> 
> Considering the current way of having separate high level commands for
> changing block parameters (block_set_hostcache, and may be block_set_flush
> in furture), these dynamic requests will be sequential. So wouldn't it 
> be better to
> avoid re-opening of image if possible for individual flag change request 
> that comes in?
>> Actually, once we know what we really want (I haven't seen many comments
>> on the BDRVReopenState suggestion yet), it should be pretty easy to
>> implement.
>>
>> Kevin
>>   
> Will work on to get an RFC patch with this proposed BDRVReopenState to 
> get more
> inputs.

Are you still going to prepare an RFC patch implementing
bdrv_reopen_prepare/commit/abort using a BDRVReopenState? Or has it even
been posted and I just missed it?

Kevin