Re: [Qemu-devel] [PATCH] sheepdog: implement direct write semantics

[Kevin Wolf]

Am 11.01.2013 08:52, schrieb MORITA Kazutaka:
> At Thu, 10 Jan 2013 13:38:16 +0800,
> Liu Yuan wrote:
>>
>> On 01/09/2013 11:10 PM, Paolo Bonzini wrote:
>>> Il 09/01/2013 14:04, Liu Yuan ha scritto:
>>>>>>   2 The upper layer software which relies on the 'cache=xxx' to choose
>>>>>> cache mode will fail its assumption against new QEMU.
>>>>>
>>>>> Which assumptions do you mean? As far as I can say the behaviour hasn't
>>>>> changed, except possibly for the performance.
>>>>
>>>> When users set 'cache=writethrough' to export only a writethrough cache
>>>> to Guest, but with new QEMU, it will actually get a writeback cache as
>>>> default.
>>>
>>> They get a writeback cache implementation-wise, but they get a
>>> writethrough cache safety-wise.  How the cache is implemented doesn't
>>> matter, as long as it "looks like" a writethrough cache.
>>>
>>
>>> In fact, consider a local disk that doesn't support FUA.  In old QEMU,
>>> images used to be opened with O_DSYNC and that splits each write into
>>> WRITE+FLUSH, just like new QEMU.  All that changes is _where_ the
>>> flushes are created.  Old QEMU changes it in the kernel, new QEMU
>>> changes it in userspace.
>>>
>>>> We don't need to communicate to the guest. I think 'cache=xxx' means
>>>> what kind of cache the users *expect* to export to Guest OS. So if
>>>> cache=writethrough set, Guest OS couldn't turn it to writeback cache
>>>> magically. This is like I bought a disk with 'writethrough' cache
>>>> built-in, I didn't expect that it turned to be a disk with writeback
>>>> cache under the hood which could possible lose data when power outage
>>>> happened.
>>>
>>> It's not by magic.  It's by explicitly requesting the disk to do this.
>>>
>>> Perhaps it's a bug that the cache mode is not reset when the machine is
>>> reset.  I haven't checked that, but it would be a valid complaint.
>>>
>>
>> Ah I didn't get the current implementation right. I tried the 3.7 kernel
>> and it works as expected (cache=writethrough result in a 'writethrough'
>> cache in the guest).
>>
>> It looks fine to me to emulate writethrough as writeback + flush, since
>> the profermance drop isn't big, though sheepdog itself support true
>> writethrough cache (no flush).
> 
> Can we drop the SD_FLAG_CMD_CACHE flag from sheepdog write requests
> when bdrv_enable_write_cache() is false?  Then the requests behave
> like FUA writes and we can safely omit succeeding flush requests.

First we would need to make sure that on a writeback -> writethrough
switch a flush happens. But once this is implemented, I think your
suggestion would work well.

Kevin