On Sun, Nov 26, 2017 at 09:01:09PM +0200, Jarkko Sakkinen wrote:
> On Sun, Nov 26, 2017 at 08:46:27PM +0200, Jarkko Sakkinen wrote:
> > On Sun, Nov 26, 2017 at 07:33:56PM +0200, Jarkko Sakkinen wrote:
> > > > +struct sgx_encl_page {
> > > > + unsigned long addr;
> > > > + unsigned int flags;
> > > > + void *epc_page;
> > > > + struct sgx_va_page *va_page;
> > > > + unsigned int va_offset;
> > > > + struct list_head list;
> > > > +};
> > >
> > > This can be without major streches hoops packed way more. There are 12
> > > bits free for use in the addr field. I can use low three bits for flags
> > > and upper nine bits for va_offset.
> > >
> > > I think I can also quite easily union epc_page and va_page fields.
> > >
> > > With these changes the struct would shrink to:
> > >
> > > struct sgx_encl_page {
> > > unsigned long addr;
> > > union {
> > > void *epc_page;
> > > struct sgx_va_page *va_page;
> > > };
> > > struct list_head list;
> > > };
> >
> > Once I started to thinking more about this I realized that this can be
> > shrunk even further. I can drop the list field and use a tag for loaded
> > pages in the radix tree where these are stored. The end result would be
> >
> > struct sgx_encl_page {
> > unsigned long addr;
> > union {
> > void *epc_page;
> > struct sgx_va_page *va_page;
> > };
> > };
>
> With this last size optimization there is an issue that it would be nice
> to select the pages to be swapped out in FIFO order. That kind of means
> that tag would not work quite well.
>
> I could try an approach where I have a pointer array to present a queue
> inside struct sgx_encl for loaded pages and in that way get rid of the
> 'list' field.
I did other optimizations except taking the list field away so it's now
a 32 byte structure. I guess sufficient for the first mainline version.
/Jarkko
