Re: [Qemu-devel] [PATCH v7 4/9] block: treat BDRV_REQ_ALLOCATE as serialising

[Max Reitz]

On 2018-01-30 13:36, Anton Nefedov wrote:
> 
> 
> On 29/1/2018 10:48 PM, Max Reitz wrote:
>> On 2018-01-18 18:49, Anton Nefedov wrote:
>>> The idea is that ALLOCATE requests may overlap with other requests.
>>> Reuse the existing block layer infrastructure for serialising requests.
>>> Use the following approach:
>>>    - mark ALLOCATE serialising, so subsequent requests to the area wait
>>>    - ALLOCATE request itself must never wait if another request is in
>>> flight
>>>      already. Return EAGAIN, let the caller reconsider.
>>>
>>> Signed-off-by: Anton Nefedov <address@hidden>
>>> Reviewed-by: Eric Blake <address@hidden>
>>> ---
>>>   block/io.c | 27 +++++++++++++++++++--------
>>>   1 file changed, 19 insertions(+), 8 deletions(-)
>>
>> The basic principle looks good to me.
>>
>>> diff --git a/block/io.c b/block/io.c
>>> index cf2f84c..4b0d34f 100644
>>> --- a/block/io.c
>>> +++ b/block/io.c
>>
>> [...]
>>
>>> @@ -1717,7 +1728,7 @@ int coroutine_fn bdrv_co_pwritev(BdrvChild *child,
>>>           struct iovec head_iov;
>>>             mark_request_serialising(&req, align);
>>> -        wait_serialising_requests(&req);
>>> +        wait_serialising_requests(&req, false);
>>
>> What if someone calls bdrv_co_pwritev() with BDRV_REQ_ZERO_WRITE |
>> BDRV_REQ_ALLOCATE?  
> 
> Either
> 
>     assert(!(qiov && (flags & BDRV_REQ_ALLOCATE)));
> 
> will fail or bdrv_co_do_zero_pwritev() will be used.

Ah, right, I didn't see that condition there.

>> .. Then this should do exactly the same as
>> bdrv_co_do_zero_pwritev(), which it currently does not -- besides this
>> serialization, this includes returning -ENOTSUP if there is a head or
>> tail to write.
>>
> 
> Another question is if that assertion is ok.
> In other words: should (qiov!=NULL && REQ_ALLOCATE) be a valid case?
> e.g. with qiov filled with zeroes?

I think it's OK to leave the assertion that way.  But maybe move it
down, under the bdrv_co_do_zero_pwritev() call (and then omit the qiov
!= NULL, because that's guaranteed then)?

(But maybe not everyone's as blind as me.)

> I'd rather document that not supported (and leave the assertion).
> 
> Actually, even (qiov!=NULL && REQ_ZERO_WRITE) looks kind of
> unsupported/broken? Alignment code in bdrv_co_pwritev() zeroes out the
> head and tail by passing the flag down bdrv_aligned_pwritev()

Yes.  Maybe we should have an assertion that you aren't allowed to pass
a qiov with REQ_ZERO_WRITE...

Max

signature.asc
Description: OpenPGP digital signature