Re: [Qemu-devel] [PATCHv2 07/20] block: honour alignment and limit in bdrv_co_do_write_zeroes

[Peter Lieven]

Am 19.09.2013 um 22:07 schrieb Eric Blake <address@hidden>:

> On 09/17/2013 07:48 AM, Peter Lieven wrote:
>> Signed-off-by: Peter Lieven <address@hidden>
>> ---
>> block.c |   57 +++++++++++++++++++++++++++++++++++++++++----------------
>> 1 file changed, 41 insertions(+), 16 deletions(-)
>> 
>> diff --git a/block.c b/block.c
>> index 177720e..74ec342 100644
>> --- a/block.c
>> +++ b/block.c
>> @@ -2660,28 +2660,53 @@ static int coroutine_fn 
>> bdrv_co_do_write_zeroes(BlockDriverState *bs,
>>     BlockDriver *drv = bs->drv;
>>     QEMUIOVector qiov;
>>     struct iovec iov;
>> -    int ret;
>> +    int ret = 0;
>> 
>> -    /* TODO Emulate only part of misaligned requests instead of letting 
>> block
>> -     * drivers return -ENOTSUP and emulate everything */
>> +    /* if no limit is specified in the BlockDriverState use a default
>> +     * of 32768 512-byte sectors (16 MiB) per request.
>> +     */
>> +    int max_write_zeroes = bs->max_write_zeroes ? bs->max_write_zeroes : 
>> 32768;
> 
> Worth having a named constant instead of a magic number?

Its the only place where it is used. I can do that, but I would keep it in 
front of the
function rather than at the top of block.h or even in block_int.h

> 
>> +    while (nb_sectors > 0 && !ret) {
> 
>> +
>> +        if (ret == -ENOTSUP) {
>> +            /* Fall back to bounce buffer if write zeroes is unsupported */
>> +            iov.iov_len  = nb_sectors * BDRV_SECTOR_SIZE;
>> +            iov.iov_base = qemu_blockalign(bs, iov.iov_len);
>> +            memset(iov.iov_base, 0, iov.iov_len);
> 
> This allocates, clears, and frees iov.iov_len bytes of 0 every iteration
> through the loop.  Can you hoist that so you only allocate the bounce
> buffer once, and then clean it up after the loop completes?

I can do that, but in this case, I have to allocate max_write_zeroes * 
BDRV_SECTOR_SIZE
bytes because the iov_len might not be maximal in the first iteration due to 
the alignment
logic.

Peter