Re: physical memory limit of 64-bit linux

[Ingo Molnar]

* Rafael J. Wysocki <rjw@xxxxxxx> wrote:

> On Tuesday, 16 of December 2008, Martin Steigerwald wrote:
> > 
> > Hi!
> > 
> > What is the physical memory limit for 64-bit Linux? I read about 40 
> > bit address bus for AMD Athlon X2 (1 TB) and 48 bit for Barcelona X4 
> > (256 TB).
> > 
> > Is 64-bit linux able to use that amount - provided that one would 
> > manage to build it into a machine? Or does it have a lower limit?
> > 
> > Looking into the Google crystal ball gives unclear pictures... I tend 
> > to assume that Linux would handle it, but I am not sure.
> 
> IIRC, the current maximal virtual memory space size of the kernel on 
> x86_64 is 2^46.

Almost: the real current upstream kernel hard memory limit on x86-64 is 44 
bits, i.e. 16 TB.

There's a couple of limits to consider here.

Firstly, there's the architectural limit imposed by the CPU - that is 48 
bits, 256 TB. That is the full virtual memory range that x86-64 CPUs are 
able to address: non-canonical addresses outside that range create an 
exception.

I.e. valid addresses on x86-64 are in the range of:

        [ 0xffff800000000000...0x00007fffffffffff ]

Which is minus 128 TB to plus 128 TB.

Traditionally (and because it's practical) that max range is split in two: 
negative addresses to kernel-space-only addresses [the same on all tasks], 
positive addresses to user-space addresses [unique to each process MM].

The kernel starts at minus 122 TB, far far down, to take maximum advantage 
of the negative range:

  arch/x86/include/asm/page_64.h:
  #define __PAGE_OFFSET           _AC(0xffff880000000000, UL)

(and we start with an 8 TB empty-mapped hole range. )

That is where all physical memory is mapped to, linearly. Then we have a 
64 TB limit imposed on the maximum size of this linear kernel memory 
range:

  #define MAXMEM           _AC(0x00003fffffffffff, UL)

that is sized a bit optimistically - it ends at ffffc7ffffffffff, which 
overlaps by 2TB into the vmalloc area, which starts at:

  #define VMALLOC_START    _AC(0xffffc20000000000, UL)

We want to set MAXMEM to VMALLOC_START-hole instead, where the hole is say 
0x20000000000. (2 TB)

This problem is academic because there are no such systems in existence, 
and because we have another limit on the size of physical memory:

  arch/x86/include/asm/sparsemem.h:
  # define MAX_PHYSMEM_BITS	44

So in reality MAXMEM should be limited to the max sparsemem-covered 
physical memory range, via the patch below.

In terms of future extensibility:

phase 1)  We could go to 45 bits (32 TB) via a twoliner patch,
          should the need arise

phase 2)  We can then go to 46 bits (64 TB) with small changes too - by 
          moving the vmalloc area up a notch and moving the followon 
          dynamic kernel mappings areas too.

phase 3)  We could also go close to 47 bits: with various more invasive
          movings of VMALLOC and rest upwards, and other considerations
          such as the elimination of the generous start of 8 TB hole at
          __PAGE_OFFSET - i.e. moving __PAGE_OFFSET straight down to
          minus 128 TB. 120 TB would be doable.

phase 4)  If the 48 bits limit is ever lifted on the CPU side, we can move
          __PAGE_OFFSET down. This is actually less invasive than phase 
          3), because moving __PAGE_OFFSET is relatively easy. The far
          more invasive change would be the necessary changes to the
          virtual memory code: the current 4-level paging has a 256 TB 
          limit which comes from the 512*512*512*512*4K split of 
          pgd/pud/pmd/pte entries. Either PGDIR_SHIFT would have to be 
          increased, moving the root pgtable's size from 4K to 8K or more, 
          or another pgdir level would have to be introduced (which is 
          even more intrusive and much less likely to be implemented by hw
          makers).

	Ingo

-------------------->