Neal> The two features that are (maybe) disabled in the -Pn
Neal> versions are:
Neal> 1) SPE[0] must always be SET in ICSR (Ibox csr)
Neal> 2) SP[0] must always be SET in the MCSR (Mbox CSR).
Neal> These features control one of the superpage mappings. It
Neal> is a static configuration that is performed during
Neal> initialisation.
OK, that wouldn't be too bad. It would mean that the top 2GB in the
virtual address space could not be mapped through the MMU because it
would identity map physical memory. I presume this feature is used in
NT to deal efficiently with (kernel) 32-bit pointers that have bit 31
set---those pointers may get sign-extended to 64 bit. If I recall
correctly, Jim's 32-bit port of Linux could have made use of this as
well.
Linux *does* use the virtual address space starting at
0xfffffe0000000000 for vmalloc'd memory (e.g., modules), but even with
the top 2GB gone, there is still 2046 GB between that address and the
starting address of the top 2GB (0xffffffff80000000). By the time we
have *that* much physical RAM, the battle of the titans will be over
(well, maybe not, but even the 21264's surely will be antique relics
by then... :-) So, not being able to turn off the additional mapping
shouldn't be a problem for Linux.
Neal> My job is to provide factory apps. support for customers
Neal> who design with Digital's silicon so I guess, on this one, I
Neal> speak for Digital.
Really appreciate your making this information available. Since Linux
is going to be around for a while, it is helpful to know what DEC
might be up to so we can avoid silly mistakes further down the road.
And anything that helps Linux is probably fine by the people on this
mailing-list... :-)
--david
Copyright © 1995-1997 Red Hat Software. Legal notices