Re: [Qemu-devel] block: regression: savevm/delvm too slow

[Kevin Wolf]

Am 21.06.2011 19:29, schrieb Luiz Capitulino:
> I'm getting the following times when doing a savevm and delvm in current
> HEAD eb47d7c5d (time in minutes, each time corresponds to a savevm/delvm run):
> 
>  savevm: 5:28m, 11:00m, 11:10m
>  delvm: 4:30m, 4:40m, > 15m
> 
> Now, trying with qemu 0.13.0 I get:
> 
>  savevm: < 1:00m, 4:00m, 4:34m
>  delvm: few seconds for all cases
> 
> Yes, you read it correctly, I tried with 0.13.0 because 0.14.0 also has the
> bug. This is the pattern I see when I run strace against HEAD while running
> the savevm command:
> 
>  pwrite(7, 
> "\0\1\0\1\0\1\0\1\0\2\0\2\0\2\0\2\0\2\0\2\0\2\0\2\0\2\0\2\0\2\0\2"..., 65536, 
> 196608) = 65536
>  fdatasync(7)                            = 0
> 
> Ie. a fdatasync() follows every single pwrite(). Something similar also 
> happens
> with delvm. I don't see this pattern with 0.13.0.
> 
> The good news is that I've tracked it down and Mr. git bisect says that:
> 
> 29c1a7301af752de6721e031d31faa48887204bd is the first bad commit
> commit 29c1a7301af752de6721e031d31faa48887204bd
> Author: Kevin Wolf <address@hidden>
> Date:   Mon Jan 10 17:17:28 2011 +0100
> 
>     qcow2: Use QcowCache
>     
>     Use the new functions of qcow2-cache.c for everything that works on 
> refcount
>     block and L2 tables.
>     
>     Signed-off-by: Kevin Wolf <address@hidden>
> 
> :040000 040000 83e364185d37845bb27f1dccd1249d14cc7a9a1e 
> 0c91964a52b5869333d4fb2cb0fa83104151359e M    block

Let me guess... You're using cache=writethrough?

Previously, qcow2_update_snapshot_refcount() had implemented its own
kind of writeback cache that would be used even with cache=writethrough.
Now we're using the generic Qcow2Cache, which implements a write-through
behaviour for cache=writethrough.

We could fix this by temporarily switching the cache to writeback mode.

Kevin