On Tue, Jan 15, 2019 at 09:54:00AM -0700, Jason Gunthorpe wrote:
> On Tue, Jan 15, 2019 at 02:25:11PM +0200, Leon Romanovsky wrote:
> > On Mon, Jan 14, 2019 at 08:37:55PM +0000, Jason Gunthorpe wrote:
> > > On Mon, Jan 14, 2019 at 04:18:22PM +0200, Leon Romanovsky wrote:
> > > > From: Leon Romanovsky <leonro@xxxxxxxxxxxx>
> > > >
> > > > Conversion to XArray allows us to mark specific entry with custom flags,
> > > > in our case we need to distinguish between user vs. kernel entries, so drop
> > > > custom boolean flag used before in favour of XAarray mark feature.
> > >
> > > Marks are mainly helpful in the xarray if they are used as part of the
> > > iteration with xa_for_each as they significantly reduce the overhead..
> > >
> > > > +bool rdma_is_kernel_res(struct rdma_restrack_entry *res)
> > > > +{
> > > > + struct ib_device *dev = res_to_dev(res);
> > > > + struct xarray *xa;
> > > > +
> > > > + xa = rdma_dev_to_xa(dev, res->type);
> > > > +
> > > > + return !xa_get_mark(xa, res_to_id(res), RES_USER_ENTRY);
> > > > +}
> > > > +EXPORT_SYMBOL(rdma_is_kernel_res);
> > >
> > > This is now O(log(n)) to save an irrelevant amount of memory.
> >
> > Yes, now it takes 3 bits which anyway were allocated as part of xa_state
> > handling. Before we had O(k) in memory footprint, while "k" is number of all
> > objects and "n" is number of valid objects only.
>
> If memory is the issue here the optimize the packing of the restrack
> entry, I don't see a good reason to use marks for this..
I see using marks as more convenient API, instead of using some bool value.
My intention is to remove duplicated fields and "user" is one of them.
Thanks
>
> Jason
Attachment:
signature.asc
Description: PGP signature
