Re: [PATCH for-4.2 1/4] Revert "qcow2: skip writing zero buffers to empty COW areas"

[Max Reitz]

On 01.11.19 13:40, Eric Blake wrote:
> On 11/1/19 11:00 AM, Max Reitz wrote:
>> This reverts commit c8bb23cbdbe32f5c326365e0a82e1b0e68cdcd8a.
>>
>> This commit causes fundamental performance problems on XFS (because
>> fallocate() stalls the AIO pipeline), and as such it is not clear that
>> we should unconditionally enable this behavior.
>>
>> We expect subclusters to alleviate the performance penalty of small
>> writes to newly allocated clusters, so when we get them, the originally
>> intended performance gain may actually no longer be significant.
>>
>> If we want to reintroduce something similar to c8bb23cbdbe, it will
>> require extensive benchmarking on various systems with subclusters
>> enabled.
>>
>> Cc: address@hidden
>> Signed-off-by: Max Reitz <address@hidden>
>> ---
> 
>> +++ b/qapi/block-core.json
>> @@ -3304,8 +3304,6 @@
>>   #
>>   # @cor_write: a write due to copy-on-read (since 2.11)
>>   #
>> -# @cluster_alloc_space: an allocation of file space for a cluster
>> (since 4.1)
>> -#
>>   # @none: triggers once at creation of the blkdebug node (since 4.1)
> 
> Deleting released qapi is not backwards-compatible.

Right. :-/

I’ll just not make changes to the QAPI schema.  It doesn’t hurt to leave
this in.

Max

> However, given that
> the only known user of this interface is debug testing via iotests, I'm
> not too concerned that we would be impacting any external users.

signature.asc
Description: OpenPGP digital signature