Re: [PATCH, RFC] virtio_blk: add cache flush command

[Anthony Liguori <anthony@codemonkey.ws>]

Christoph Hellwig wrote:
> On Mon, May 11, 2009 at 09:51:40AM -0500, Anthony Liguori wrote:
>   
>> What typically triggers a flush operation?
>>     
>
> fsync.
>
>   
>> I would assume an fsync would, but would a flush happen after every 
>> O_DIRECT write?
>>     
>
> Right now it doesn't, but it probably should.
>   

So then with cache=writeback, fsync behaves itself but O_DIRECT writes 
do not.

This seems like a really undesirable combination of behavior from a 
guest integrity point of view.  It makes me wonder if it's really 
useful.  I think that any serious user would have to continue using 
cache=writethrough.  Is there a path that would ever allow someone who 
cares about their data to use cache=writeback instead of cache=writethrough?

>> If the backend implementation of T_FLUSH is fsync, I would think that 
>> this would result in rather poor performance for O_DIRECT operations in 
>> the guest.
>>     
>
> Right now it's fsync.  By the time I'll submit the backend change it
> will still be fsync, but at least called from the posix-aio-compat
> thread pool.
>   

fsync is pretty crappy on ext3 default configs.  I'm concerned that this 
could be considered a DoS by a malicious guest.  If it sat in a T_FLUSH 
loop, it would potentially bring your system to a crawl, no?

Regards,

Anthony Liguori