Re: [Qemu-devel] [RFC] block-queue: Delay and batch metadata writes

[Kevin Wolf <kwolf@redhat.com>]

Am 20.09.2010 16:31, schrieb Anthony Liguori:
> On 09/20/2010 08:56 AM, Kevin Wolf wrote:
>> I won't get this ready until I leave for vacation on Wednesday, so I thought I
>> could just as well post it as an RFC in this state.
>>
>> With this patch applied, qcow2 doesn't directly access the image file any more
>> for metadata, but rather goes through the newly introduced blkqueue. Write
>> and sync requests are queued there and executed in a separate worker thread.
>> Reads consider the contents of the queue before accessing the the image file.
>>
>> What makes this interesting is that we can delay syncs and if multiple syncs
>> occur, we can merge them into one bdrv_flush.
>>
>> A typical sequence in qcow2 (simple cluster allocation) looks like this:
>>
>> 1. Update refcount table
>> 2. bdrv_flush
>> 3. Update L2 entry
>>    
> 
> Let's expand it a bit more:
> 
> 1. Update refcount table
> 2. bdrv_flush
> 3. Update L2 entry
> 4. Write data to disk
> 5. Report write complete
> 
> I'm struggling to understand how a thread helps out.

This sequence becomes:

1. Update refcount table
2. Write data to disk
3. Report write complete

And only later:

4. Update L2 entry
5. bdrv_flush (possibly merged with other flushes)

> If you run 1-3 in a thread, you need to inject a barrier between steps 3 
> and 5 or you'll report the write complete before writing the metadata 
> out.  You can't delay completing step 3 until a guest requests a flush.  
> If you do, then you're implementing a writeback cache for metadata.

Yeah, if you like to call it that, that's probably an accurate description.

> If you're comfortable with a writeback cache for metadata, then you 
> should also be comfortable with a writeback cache for data in which 
> case, cache=writeback is the answer.

Well, there is a difference: We don't pollute the host page cache with
guest data and we don't get a virtual "disk cache" as big as the host
RAM, but only a very limited queue of metadata.

Basically, in qemu we have three different types of caching:

1. O_DSYNC, everything is always synced without any explicit request.
   This is cache=writethrough.

2. Nothing is ever synced. This is cache=unsafe.

3. We present a writeback disk cache to the guest and the guest needs
   to explicitly flush to gets its data safe on disk. This is
   cache=writeback and cache=none.

So they are actually very similar, the difference is only if to use
O_DIRECT or not. In principle, regarding the integrity requirements
there is already no difference between cache=none and cache=writeback today.

We're still lacking modes for O_DSYNC | O_DIRECT and unsafe | O_DIRECT,
but they are entirely possible, because it's two different dimensions.
(And I think Christoph was planning to actually make it two independent
options)

You have a point in that we need to disable the queueing for
cache=writethrough. I'm aware of that, but forgot to mention it in the
todo list.

> If it's a matter of batching, batching can't occur if you have a barrier 
> between steps 3 and 5.  The only way you can get batching is by doing a 
> writeback cache for the metadata such that you can complete your request 
> before the metadata is written.
> 
> Am I misunderstanding the idea?

No, I think you understand it right, but maybe you were not completely
aware that cache=none doesn't mean writethrough.

Kevin