Re: [Qemu-devel] [PATCH] block: Avoid unecessary drv->bdrv_getlength() calls

[Fam Zheng <famz@redhat.com>]

On 11/04/2013 07:18 PM, Kevin Wolf wrote:
> Am 04.11.2013 um 08:24 hat Fam Zheng geschrieben:
>>
>> On Tue 29 Oct 2013 08:12:38 PM CST, Kevin Wolf wrote:
>>> Am 29.10.2013 um 13:02 hat Paolo Bonzini geschrieben:
>>>> Il 29/10/2013 12:35, Kevin Wolf ha scritto:
>>>>> The block layer generally keeps the size of an image cached in
>>>>> bs->total_sectors so that it doesn't have to perform expensive
>>>>> operations to get the size whenever it needs it.
>>>>>
>>>>> This doesn't work however when using a backend that can change its size
>>>>> without qemu being aware of it, i.e. passthrough of removable media like
>>>>> CD-ROMs or floppy disks. For this reason, the caching is disabled when a
>>>>> removable device is used.
>>>>>
>>>>> It is obvious that checking whether the _guest_ device has removable
>>>>> media isn't the right thing to do when we want to know whether the size
>>>>> of the host backend can change. To make things worse, non-top-level
>>>>> BlockDriverStates never have any device attached, which makes qemu
>>>>> assume they are removable, so drv->bdrv_getlength() is always called on
>>>>> the protocol layer. In the case of raw-posix, this causes unnecessary
>>>>> lseek() system calls, which turned out to be rather expensive.
>>>>>
>>>>> This patch completely changes the logic and disables bs->total_sectors
>>>>> caching only for certain block driver types, for which a size change is
>>>>> expected: host_cdrom and host_floppy; also the raw format in case it
>>>>> sits on top of one of these protocols, but in the common case the nested
>>>>> bdrv_getlength() call on the protocol driver will use the cache again
>>>>> and avoid an expensive drv->bdrv_getlength() call.
>>>>>
>>>>> Signed-off-by: Kevin Wolf <kwolf@redhat.com>
>>>> raw-win32.c probably needs to have a .has_variable_length=true in
>>>> bdrv_host_device.  Apart from that,
>>>>
>>>> Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
>>> Thanks, good catch. I've added this now.
>> This breaks VMDK because it can't read description file: buffer is
>> empty in bdrv_probe, when the file size is <512 bytes. (for
>> raw-posix).
> Perhaps we should replace bs->total_sectors with bs->total_byts one
> day... For now, how about introducing a bdrv_getlength_bytes() that
> always calls the driver and never converts bytes to sectors?
I think it will work. And is it correct to round up total_bytes to 
BDRV_SECTOR_SIZE before storing to total_sectors? From my view it's 
better than losing some bytes by rounding down.

Fam