[pve-devel] [PATCH qemu-server] add discard_granularity to 4M for rbd storage

[Thomas Lamprecht]

On 6/28/18 2:56 PM, Alwin Antreich wrote:
> On Thu, Jun 28, 2018 at 10:08:45AM +0200, Alexandre Derumier wrote:
>> when we have snapshots on rbd and do a trim, the space is increasing
>> http://tracker.ceph.com/issues/18352
>>
>> we need to trim a full object (4MB by default), to be able to free space.
>>
>> test:
>>
>> without discard_granularity
>> ---------------------------
>> vm-107-disk-1		                                       20480M   2500M
>> vm-107-disk-1                                                  20480M   2500M
>>
>> vm-107-disk-1 at snap1                                            20480M   2500M
>> vm-107-disk-1                                                  20480M  90112k
>>
>> vm-107-disk-1 at snap1                                            20480M   2500M
>> vm-107-disk-1                                                  20480M   1020M
>>
>> with discard_granularity=4M
>> ---------------------------
>> vm-107-disk-1		                                       20480M   2500M
>> vm-107-disk-1                                                  20480M   2500M
>>
>> vm-107-disk-1 at snap1                                            20480M   2500M
>> vm-107-disk-1                                                  20480M  90112k
>>
>> vm-107-disk-1 at snap1                                            20480M   2500M
>> vm-107-disk-1                                                  20480M    144M
>> ---
>>  PVE/QemuServer.pm | 5 +++++
>>  1 file changed, 5 insertions(+)
>>
>> diff --git a/PVE/QemuServer.pm b/PVE/QemuServer.pm
>> index 6a355f8..fd9754c 100644
>> --- a/PVE/QemuServer.pm
>> +++ b/PVE/QemuServer.pm
>> @@ -1695,6 +1695,11 @@ sub print_drivedevice_full {
>>  	$device .= ",serial=$serial";
>>      }
>>  
>> +    my $volid = $drive->{file};
>> +    if($volid && $drive->{discard}) {
>> +        my $storage_name = PVE::Storage::parse_volume_id($volid);
>> +        $device .= ",discard_granularity=4194304" if $storecfg->{ids}->{$storage_name}->{type} eq 'rbd';
>> +    }
> In my opinion it would be better to have Qemu figure it out automagically or
> use the discard config option to add the granularity.
> 
> As example: discard=on / discard=4194304 (is on with granularity).
> 
> This way it is configurable per disk image and can be set according to the
> needs of the storage (eg. zvol; 8K).
> 
> For Ceph, the object size can be set when an image is created (--object-size).
> 

As long as we know which storage has which size, and that stays fixed over
the image lifetime, such a fix is OK, IMO. He  especially checked for rbd so
zvol stays the same.

If we deem that no-good and decide to add it as integer (in 1K or 4K multiples
please) to the disk config entry, the questions is: what do with existing
images?