Re: [Qemu-devel] [PATCH] qcow2: Add full image preallocation option

[Kevin Wolf <kwolf@redhat.com>]

Am 27.01.2011 18:50, schrieb Anthony Liguori:
> On 01/27/2011 09:58 AM, Daniel P. Berrange wrote:
>> On Thu, Jan 27, 2011 at 04:52:14PM +0100, Kevin Wolf wrote:
>>    
>>> This adds a preallocation=full mode to qcow2 image creation, which does not
>>> only allocate metadata for the whole image, but also writes zeros to it,
>>> creating a non-sparse image file.
>>>
>>> Signed-off-by: Kevin Wolf<kwolf@redhat.com>
>>> ---
>>>   block/qcow2.c |   45 ++++++++++++++++++++++++++++++++++++++++-----
>>>   1 files changed, 40 insertions(+), 5 deletions(-)
>>>
>>> diff --git a/block/qcow2.c b/block/qcow2.c
>>> index a1773e4..90cf2ca 100644
>>> --- a/block/qcow2.c
>>> +++ b/block/qcow2.c
>>> @@ -838,7 +838,15 @@ static int qcow2_change_backing_file(BlockDriverState *bs,
>>>       return qcow2_update_ext_header(bs, backing_file, backing_fmt);
>>>   }
>>>
>>> -static int preallocate(BlockDriverState *bs)
>>> +enum prealloc_mode {
>>> +    PREALLOC_OFF = 0,
>>> +    PREALLOC_METADATA,
>>> +    PREALLOC_FULL,
>>> +};
>>> +
>>> +#define IO_BUF_SIZE (2 * 1024 * 1024)
>>> +
>>> +static int preallocate(BlockDriverState *bs, enum prealloc_mode mode)
>>>   {
>>>       uint64_t nb_sectors;
>>>       uint64_t offset;
>>> @@ -846,11 +854,14 @@ static int preallocate(BlockDriverState *bs)
>>>       int ret;
>>>       QCowL2Meta meta;
>>>
>>> +    assert(mode != PREALLOC_OFF);
>>> +
>>>       nb_sectors = bdrv_getlength(bs)>>  9;
>>>       offset = 0;
>>>       QLIST_INIT(&meta.dependent_requests);
>>>       meta.cluster_offset = 0;
>>>
>>> +    /* First allocate metadata in _really_ big chunks */
>>>       while (nb_sectors) {
>>>           num = MIN(nb_sectors, INT_MAX>>  9);
>>>           ret = qcow2_alloc_cluster_offset(bs, offset, 0, num,&num,&meta);
>>> @@ -874,6 +885,28 @@ static int preallocate(BlockDriverState *bs)
>>>           offset += num<<  9;
>>>       }
>>>
>>> +    /* Then write zeros to the cluster data, if requested */
>>> +    if (mode == PREALLOC_FULL) {
>>> +        void *buf = qemu_mallocz(IO_BUF_SIZE);
>>> +
>>> +        nb_sectors = bdrv_getlength(bs)>>  BDRV_SECTOR_BITS;
>>> +        offset = 0;
>>> +
>>> +        while (nb_sectors) {
>>> +            num = MIN(nb_sectors, IO_BUF_SIZE / BDRV_SECTOR_SIZE);
>>> +            ret = bdrv_write(bs, offset>>  BDRV_SECTOR_BITS, buf, num);
>>>      
>> Is there a way you can calculate the total size of the qcow2
>> file upfront, and just use a single posix_fallocate() call to
>> do the zero-filled allocation of all the data blocks. It is
>> many orders of magnitude faster than truely writing blocks of
>> zero'd data on modern filesystems.  I guess if you're using
>> compression or encryption, we'd really have to go the slow
>> path, but for regular usage it'd be better to take a fast
>> path.
>>    
> 
> Hrm, so is the intention here to avoid sparse files or to not assume 
> zero-fill?

The primary intention (as I understood our feature request ;-)) was to
avoid sparse files. In it's current implementation you could also use it
to overwrite any left-over data. Maybe that's a point for not having a
bdrv_fallocate like Daniel suggested, but rather a bdrv_zero_init, which
could fallocate on files and write zeros on a block device.

Kevin