On 12/7/22 6:09 AM, Yu Kuai wrote: > Hi, > > 在 2022/12/07 11:15, Ming Lei 写道: >> On Wed, Dec 07, 2022 at 10:19:08AM +0800, Yu Kuai wrote: >>> Hi, >>> >>> 在 2022/12/07 2:15, Gulam Mohamed 写道: >>>> Use ktime to change the granularity of IO accounting in block layer from >>>> milli-seconds to nano-seconds to get the proper latency values for the >>>> devices whose latency is in micro-seconds. After changing the granularity >>>> to nano-seconds the iostat command, which was showing incorrect values for >>>> %util, is now showing correct values. >>> >>> This patch didn't correct the counting of io_ticks, just make the >>> error accounting from jiffies(ms) to ns. The problem that util can be >>> smaller or larger still exist. >> >> Agree. >> >>> >>> However, I think this change make sense consider that error margin is >>> much smaller, and performance overhead should be minimum. >>> >>> Hi, Ming, how do you think? >> >> I remembered that ktime_get() has non-negligible overhead, is there any >> test data(iops/cpu utilization) when running fio or t/io_uring on >> null_blk with this patch? > > Yes, testing with null_blk is necessary, we don't want any performance > regression. null_blk is fine as a substitute, but I'd much rather run this on my test bench with actual IO and devices. > BTW, I thought it's fine because it's already used for tracking io > latency. Reading a nsec timestamp is a LOT more expensive than reading jiffies, which is essentially free. If you look at the amount of work that's gone into minimizing ktime_get() for the fast path in the IO stack, then that's a testament to that. So that's a very bad assumption, and definitely wrong. -- Jens Axboe
