[Drbd-dev] [PATCH RFC 1/5] block: return ns precision from disk_start_io_acct

[Guoqing Jiang]

Hi Ming,

On 7/8/20 3:27 PM, Ming Lei wrote:
> On Wed, Jul 08, 2020 at 09:58:15AM +0200, Guoqing Jiang wrote:
>> Currently the duration accounting of bio based driver is converted from
>> jiffies to ns, means it could be less accurate as request based driver.
>>
>> So let disk_start_io_acct return from ns precision, instead of convert
>> jiffies to ns in disk_end_io_acct.
>>
>> Cc: Philipp Reisner <philipp.reisner at linbit.com>
>> Cc: Lars Ellenberg <lars.ellenberg at linbit.com>
>> Cc: drbd-dev at lists.linbit.com
>> Signed-off-by: Guoqing Jiang <guoqing.jiang at cloud.ionos.com>
>> ---
>>   block/blk-core.c | 7 ++++---
>>   1 file changed, 4 insertions(+), 3 deletions(-)
>>
>> diff --git a/block/blk-core.c b/block/blk-core.c
>> index d9d632639bd1..0e806a8c62fb 100644
>> --- a/block/blk-core.c
>> +++ b/block/blk-core.c
>> @@ -1466,6 +1466,7 @@ unsigned long disk_start_io_acct(struct gendisk *disk, unsigned int sectors,
>>   	struct hd_struct *part = &disk->part0;
>>   	const int sgrp = op_stat_group(op);
>>   	unsigned long now = READ_ONCE(jiffies);
>> +	unsigned long start_ns = ktime_get_ns();
>>   
>>   	part_stat_lock();
>>   	update_io_ticks(part, now, false);
>> @@ -1474,7 +1475,7 @@ unsigned long disk_start_io_acct(struct gendisk *disk, unsigned int sectors,
>>   	part_stat_local_inc(part, in_flight[op_is_write(op)]);
>>   	part_stat_unlock();
>>   
>> -	return now;
>> +	return start_ns;
>>   }
>>   EXPORT_SYMBOL(disk_start_io_acct);
>>   
>> @@ -1484,11 +1485,11 @@ void disk_end_io_acct(struct gendisk *disk, unsigned int op,
>>   	struct hd_struct *part = &disk->part0;
>>   	const int sgrp = op_stat_group(op);
>>   	unsigned long now = READ_ONCE(jiffies);
>> -	unsigned long duration = now - start_time;
>> +	unsigned long duration = ktime_get_ns() - start_time;
>>   
>>   	part_stat_lock();
>>   	update_io_ticks(part, now, true);
>> -	part_stat_add(part, nsecs[sgrp], jiffies_to_nsecs(duration));
>> +	part_stat_add(part, nsecs[sgrp], duration);
>>   	part_stat_local_dec(part, in_flight[op_is_write(op)]);
>>   	part_stat_unlock();
> Hi Guoqing,
>
> Cost of ktime_get_ns() can be observed as not cheap in high IOPS device,

Could you share some links about it? Thanks.

> so not sure the conversion is good. Also could you share what benefit we can
> get with this change?

Without the conversion, we have to track io latency with jiffies in 4th 
patch.
Then with HZ=100, some rows (such as 1ms, 2ms and 4ms) in that table
don't make sense.

Thanks,
Guoqing