Re: [RFD] CAT user space interface revisited

[Marcelo Tosatti]

On Tue, Nov 24, 2015 at 07:25:43PM -0200, Marcelo Tosatti wrote:
> On Tue, Nov 24, 2015 at 04:27:54PM +0800, Chao Peng wrote:
> > On Wed, Nov 18, 2015 at 10:01:54PM -0200, Marcelo Tosatti wrote:
> > > > 	tglx
> > > 
> > > Again: you don't need to look into the MSR table and relate it 
> > > to tasks if you store the data as:
> > > 
> > > 	task group 1 = {
> > > 			reservation-1 = {size = 80Kb, type = data, socketmask = 0xffff},
> > > 			reservation-2 = {size = 100Kb, type = code, socketmask = 0xffff}
> > > 	}
> > > 	
> > > 	task group 2 = {
> > > 			reservation-1 = {size = 80Kb, type = data, socketmask = 0xffff},
> > > 			reservation-3 = {size = 200Kb, type = code, socketmask = 0xffff}
> > > 	}
> > > 
> > > Task group 1 and task group 2 share reservation-1.
> > 
> > Because there is only size but not CBM position info, I guess for
> > different reservations they will not overlap each other, right?
> 
> Reservation 1 is shared between task group 1 and task group 2 
> so the CBMs overlap (by 80Kb, rounded).
> 
> > Personally I like this way of exposing minimal information to userspace.
> > I can think it working well except for one concern of losing flexibility:
> > 
> > For instance, there is a box for which the full CBM is 0xfffff. After
> > cache reservation creating/freeing for a while we then have reservations:
> > 
> > reservation1: 0xf0000
> > reservation2: 0x00ff0
> > 
> > Now people want to request a reservation which size is 0xff, so how
> > will kernel do at this time? It could return just error or do some
> > moving/merging (e.g. for reservation2: 0x00ff0 => 0x0ff00) and then
> > satisfy the request. But I don't know if the moving/merging will cause
> > delay for tasks that is using it.
> 
> Right, i was thinking of adding a "force" parameter. 
> 
> So, default behaviour of attach: do not merge.
> "force" behaviour of attach: move reservations around and merge if
> necessary.

To make the decision userspace would need the know that a merge can
be performed if particular reservations can be moved (that is, the
moveable property is per-reservation, depending on whether its ok 
for the given app to cacheline fault or not).
Anyway, thats for later.






--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/