> > So I guess we won't be switching pg->phys_addr from paddr to pfn? > > Speaking of which, any plans for expanding to >32-bit (or >31-bit, if > > signed) pfns in the rest of uvm? > > That's not part of my current plans for UVM, which right now extend only as > far as breaking the back of the performance problems with builds. 31 bit > pfns get us into the terabyte range I think. I have a couple of thoughts > here, firstly SVR4 or Solaris has a pfn_t and pgcnt_t or something like that > and it would would be nice to have an analogue. Kind of surprised we didn't > inherit something like that from Mach.
FWIW, the numbers are: we break at 8TiB ram, that's where the 32 bit signed overflow happens. the other number of note: HP sell a machine with 48TiB of ram already, since most of last year if not more. another method would be to lie to UVM and provide it with >4KiB pages to manage, like VAX already does. .mrg.