On Fri, Oct 16, 2015 at 2:22 PM, Alex Elder <elder@xxxxxxxx> wrote:
> On 10/07/2015 12:00 PM, Ilya Dryomov wrote:
>> Commit 30e2bc08b2bb ("Revert "block: remove artifical max_hw_sectors
>> cap"") restored a clamp on max_sectors. It's now 2560 sectors instead
>> of 1024, but it's not good enough: we set max_hw_sectors to rbd object
>> size because we don't want object sized I/Os to be split, and the
>> default object size is 4M.
>>
>> So, set max_sectors to max_hw_sectors in rbd at queue init time.
>>
>> Signed-off-by: Ilya Dryomov <idryomov@xxxxxxxxx>
>> ---
>> drivers/block/rbd.c | 1 +
>> 1 file changed, 1 insertion(+)
>>
>> diff --git a/drivers/block/rbd.c b/drivers/block/rbd.c
>> index 05072464d25e..04e69b4df664 100644
>> --- a/drivers/block/rbd.c
>> +++ b/drivers/block/rbd.c
>> @@ -3760,6 +3760,7 @@ static int rbd_init_disk(struct rbd_device *rbd_dev)
>> /* set io sizes to object size */
>> segment_size = rbd_obj_bytes(&rbd_dev->header);
>> blk_queue_max_hw_sectors(q, segment_size / SECTOR_SIZE);
>> + q->limits.max_sectors = queue_max_hw_sectors(q);
>
> This currently is done by default by blk_queue_max_hw_sectors().
> Do you see any different behavior with this patch in place?
>
> This change seems at least harmless so if it's not too late:
Not quite. The commit mentioned above reintroduced this bit into
blk_limits_max_hw_sectors() in 4.3-rc1:
241 limits->max_sectors = min_t(unsigned int, max_hw_sectors,
242 BLK_DEF_MAX_SECTORS);
So this patch fixes a performance regression. If you spin up 4.3-rc1+
and map an rbd image without changing any settings, the best you'll see
is 1280k I/Os despite a 4096k (typically) max_hw_sectors.
Thanks,
Ilya
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
