Re: [PATCH v2 1/4] blk-iocost: charge flushes as pageless random writes

[Tejun Heo <tj@kernel.org>]

Hello,

On Sun, Sep 20, 2026 at 03:08:54PM +0800, Tao Cui wrote:
> One remaining question is whether builtin profiles should eventually carry
> measured flushiops values, or whether flushiops is expected to remain
> user-configured.  This series leaves it at zero because I don't have
> reliable measurements for the existing profiles.

The cost of a flush is more a function of what came before the flush how
long ago than anything else. The linear model is too dumb to estimate that.
You can tune this to suit a particular use case and maybe one can argue that
charging a small cost to penalize unnecessarily frequent flushes would be a
reasonable default policy, but, I don't know, what the default model can do
is pretty limited, so I'm not sure it matters that much one way or the
other.

If getting this right is actually useful, stateful BPF model is probably the
right way.

Thanks.

-- 
tejun