RE: [PATCH 3/8] scsi: ufshpb: Add region's reads counter

[Avri Altman]

> On Mon, Feb 01, 2021 at 08:17:59AM +0000, Avri Altman wrote:
> > >
> > > On Mon, Feb 01, 2021 at 07:51:19AM +0000, Avri Altman wrote:
> > > > >
> > > > > On Mon, Feb 01, 2021 at 07:12:53AM +0000, Avri Altman wrote:
> > > > > > > > +#define WORK_PENDING 0
> > > > > > > > +#define ACTIVATION_THRSHLD 4 /* 4 IOs */
> > > > > > > Rather than fixing it with macro, how about using sysfs and make it
> > > > > > > configurable?
> > > > > > Yes.
> > > > > > I will add a patch making all the logic configurable.
> > > > > > As all those are hpb-related parameters, I think module parameters
> are
> > > > > more adequate.
> > > > >
> > > > > No, this is not the 1990's, please never add new module parameters to
> > > > > drivers.  If not for the basic problem of they do not work on a
> > > > > per-device basis, but on a per-driver basis, which is what you almost
> > > > > never want.
> > > > OK.
> > > >
> > > > >
> > > > > But why would you want to change this value, why can't the driver
> "just
> > > > > work" and not need manual intervention?
> > > > It is.
> > > > But those are a knobs each vendor may want to tweak,
> > > > So it'll be optimized with its internal device's implementation.
> > > >
> > > > Tweaking the parameters, as well as the entire logic, is really an endless
> > > task.
> > > > Some logic works better for some scenarios, while falling behind on
> others.
> > >
> > > Shouldn't the hardware know how to handle this dynamically?  If not, how
> > > is a user going to know?
> > There is one "brain".
> > It is either in the device - in device mode, Or in the host - in host mode
> control.
> > The "brain" decides which region is active, thus carrying the physical address
> along with the logical -
> > minimizing context switches in the device's RAM.
> >
> > There can be up to N active regions.
> > Activation and deactivation has its overhead.
> > So basically it is a constraint-optimization problem.
> 
> So how do you solve it?  And how would you expect a user to solve it if
> the kernel can not?
> 
> You better document the heck out of these configuration options :)
Yes.  Will do.

Thanks,
Avri