1 |
Billy Holmes posted <43972B9D.1040605@××××××.net>, excerpted below, on |
2 |
Wed, 07 Dec 2005 13:36:13 -0500: |
3 |
|
4 |
> Duncan wrote: |
5 |
>> The current AMD spec says the CPUs offer 40-bit physical memory |
6 |
>> addressing, 48-bit virtual memory addressing. So, 256 TB virtual |
7 |
>> memory, but only a terabyte physical memory, in a flat-address |
8 |
>> configuration. |
9 |
> |
10 |
> thanks Duncan! that's good to know :) |
11 |
> |
12 |
> can you imagine the RAID array needed for the SWAP partition to get that |
13 |
> 256 TB of virtual memory? |
14 |
|
15 |
I'm /definitely/ not sure on this, hopefully someone else will correct me |
16 |
if I'm wrong, but I /believe/ "virtual address space" or "virtual memory" |
17 |
in this case means something other than swap. I /believe/ swap would |
18 |
still be part of the physical memory address space. |
19 |
|
20 |
/If/ I understand it correctly, the "virtual" in this case refers not to |
21 |
swap, but to the implementation where in 32-bit flat address usage, the |
22 |
address space was 4 gig, but that was divided between the kernel (using |
23 |
the upper addresses) and userland (using the lower addresses). PCI device |
24 |
addresses were traditionally mapped into the top of the upper space, 3.5-4 |
25 |
gig, but those were "virtual" addresses in the sense that it didn't mean |
26 |
the entire 4 gig was mapped -- there could be and normally were holes in |
27 |
the mapping, so folks didn't have to run a swap large enough to make up |
28 |
the difference between physical memory and 4 gig. |
29 |
|
30 |
Again, /if/ this reading is correct, the total of swap and physical memory |
31 |
cannot exceed a terabyte (40 bits addressable) in the flat memory address |
32 |
scheme, but that memory can be mapped anywhere within a 256 TB virtual |
33 |
memory space. Thus, the kernel /could/ still use the top of that virtual |
34 |
space, letting userland use the bottom of it, if the OS folks decided to |
35 |
do it that way. (In practice, I believe at least a portion of kernel |
36 |
space is still mapped much lower, for compatibility reasons. I |
37 |
/definitely/ know that the 3.5-4 GB space remains reserved for 32-bit PCI |
38 |
compatibility, because certain PCI devices can't handle anything above |
39 |
that. I know that, having read about the problems folks may have getting |
40 |
the computer to recognize all their memory, when they have 4 gig or more, |
41 |
because the PCI space at the top of the 4 gig overwrites the real memory |
42 |
at that location, if the BIOS hasn't been updated and configured to avoid |
43 |
it by mapping the otherwise overwritten physical memory addresses up above |
44 |
four gigs.) |
45 |
|
46 |
If that's an incorrect understanding, as I said, hopefully someone |
47 |
informed enough as to the true situation will correct my error. If it's |
48 |
correct, and someone informed as to the fact can verify, I'd definitely |
49 |
appreciate that, too. |
50 |
|
51 |
Anyway, in this case, it VERY MUCH depends on what one's definition of |
52 |
"virtual" is! <g> (For non-USians, this is an attempt at humor, |
53 |
being an oblique reference to the Clinton Monicagate political scandal |
54 |
here in the US a few years ago.) |
55 |
|
56 |
-- |
57 |
Duncan - List replies preferred. No HTML msgs. |
58 |
"Every nonfree program has a lord, a master -- |
59 |
and if you use the program, he is your master." Richard Stallman in |
60 |
http://www.linuxdevcenter.com/pub/a/linux/2004/12/22/rms_interview.html |
61 |
|
62 |
|
63 |
-- |
64 |
gentoo-amd64@g.o mailing list |