Re: [Qemu-devel] [Fwd: [PATCH v2] vpc: Implement bdrv_co_get_block_status()]

[Peter Lieven <pl@kamp.de>]

Am 19.02.2015 um 13:07 schrieb Kevin Wolf:
> Am 18.02.2015 um 22:03 hat Peter Lieven geschrieben:
>> Am 18.02.2015 um 21:57 schrieb Peter Lieven:
>>> This implements bdrv_co_get_block_status() for VHD images. This can
>>> significantly speed up qemu-img convert operation because only with this
>>> function implemented sparseness can be considered. (Before, converting a
>>> 1 TB empty image took several minutes for me, now it's instantaneous.)
>>>
>>> Signed-off-by: Kevin Wolf <kwolf@redhat.com>
>>> ---
>>>  block/vpc.c | 50 ++++++++++++++++++++++++++++++++++++++++++++++++--
>>>  1 file changed, 48 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/block/vpc.c b/block/vpc.c
>>> index 7fddbf0..1533b6a 100644
>>> --- a/block/vpc.c
>>> +++ b/block/vpc.c
>>> @@ -597,6 +597,51 @@ static coroutine_fn int vpc_co_write(BlockDriverState
>>> *bs, int64_t sector_num,
>>>      return ret;
>>>  }
>>>
>>> +static int64_t coroutine_fn vpc_co_get_block_status(BlockDriverState *bs,
>>> +        int64_t sector_num, int nb_sectors, int *pnum)
>>> +{
>>> +    BDRVVPCState *s = bs->opaque;
>>> +    VHDFooter *footer = (VHDFooter*) s->footer_buf;
>>> +    int64_t start, offset, next;
>>> +    bool allocated;
>>> +    int n;
>>> +
>>> +    if (be32_to_cpu(footer->type) == VHD_FIXED) {
>>> +        *pnum = nb_sectors;
>>> +        return BDRV_BLOCK_RAW | BDRV_BLOCK_OFFSET_VALID | BDRV_BLOCK_DATA |
>>> +               (sector_num << BDRV_SECTOR_BITS);
>>> +    }
>>> +
>>> +    offset = get_sector_offset(bs, sector_num, 0);
>>> +    start = offset;
>>> +    allocated = (offset != -1);
>>> +    *pnum = 0;
>>> +
>>> +    do {
>>> +        /* All sectors in a block are contiguous (without using the
>>> bitmap) */
>>> +        n = ROUND_UP(sector_num + 1, s->block_size / BDRV_SECTOR_SIZE)
>>> +          - sector_num;
>>> +        n = MIN(n, nb_sectors);
>>> +
>>> +        *pnum += n;
>>> +        sector_num += n;
>>> +        nb_sectors -= n;
>>> +        next = start + (*pnum * BDRV_SECTOR_SIZE);
>>> +
>>> +        if (nb_sectors == 0) {
>>> +            break;
>>> +        }
>>> +
>>> +        offset = get_sector_offset(bs, sector_num, 0);
>>> +    } while ((allocated && offset == next) || (!allocated && offset == -1));
>>> +
>>> +    if (allocated) {
>>> +        return BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID | start;
>>> +    } else {
>>> +        return 0;
>> Shouldn't this be
>>
>>  return BDRV_BLOCK_ZERO;
>>
>> ?
>>
>> vpc_read memsets all blocks with offset == -1 to 0x00.
> Yes, but the blocks are still unallocated, as opposed to allocated as
> zero clusters, and this is indicated by 0.

Okay, than I somehow have to fix that up in the iscsi driver. There
I tread unallocated and anchored identically. Has that changed somewhen
after the initial introduction of bdrv_get_block_status?

>
> vpc_get_info() sets bdi->unallocated_blocks_are_zero = true, so we end
> up with bdrv_co_get_block_status() returning BDRV_BLOCK_ZERO, but not
> BDRV_BLOCK_ALLOCATED (which would be set if we had BDRV_BLOCK_ZERO
> here).
>
> I'm not sure if a wrong allocated flag would cause problem currently,
> but it's definitely necessary to get right once we add support for
> differencing images (patches are on the list, pending review).
>
>> Not for this patch, but couldn't we use your new function to signifincantly speed up
>> reading of continous allocated areas in vpc_read?
> There aren't really contiguous blocks in VHD, you always have a bitmap
> in between. In some cases it might be better to read the bitmap as well
> as the two adjacent blocks and throw that buffer away in order to save
> one read request, but with relatively large block sizes of VHD it's
> probably not going to help that much.

If there is always a bitmap between 2 clusters I do not understand the loop
in your bdrv_co_get_block_status implementation for VPC?

If I understand correctly you skip over the bitmaps in the loop and report
continous sectors as allocated from start.

If there are bitmaps in between at least qemu-img map would produce wrong
output.

Peter