Re: [coreboot] 8 GB DIMMs on Nehalem (Arrandale)

[Stefan Tauner]

On Sat, 28 Jan 2017 17:01:05 +0100
Zoran Stojsavljevic  wrote:
> Hello Stefan,
>
> Let me ask you for some other stuff, since I would like to put what I wrote
> initially to hold (sleep state, for now).
>
> You wrote: *The official specs are not trustworthy IMHO and cpuid(1) and
> /proc/cpuinfo **show the same physical address width of 36 bits (which
> would indicate a **maximum of 64 GB).*
>
> Question to you: are you dealing with i686 kernel, (32 bit)? It seems to me
> that you have Nehalem which complies in IA32 with PAE HW extension, don't
> you?!
>
> This is very important -> *Enabling PAE (by setting bit 5, PAE, of the
> system register CR4) causes major changes to this scheme...*

Both CPUs have PAE support, yes, but I think I have never even tried to
boot a 32-bit kernel on them :) However, cpuid with an x86_64 kernel shows
(assumably via the cpuid op and edx):

physical address extensions            = true

But that does only indicate support not use of PAE... I don't see how
this would be relevant to memtest or the 64-bit kernels anyway.

8 GB DIMMs on Nehalem (Arrandale)

[Zoran Stojsavljevic]

> Both CPUs have PAE support, yes, but I think I have never even tried to
> boot a 32-bit kernel on them :)

This is one missing info. Because, so far, if you did not know, there are
two types of 32-bit kernels: normal 32 bit, and PAE 32 bit. But, in
contrary, x86_64 has only one type of kernel support regarding MMU.

Why? Because for i686 the following is true: Memory Management Unit (MMU)
differently are expected to emulate the three-level page tables. For
example, on the x86 without PAE enabled, only two page table levels are
available.

With non PAE 32 kernels, the PA space is limited to 32 bits, thus making
ONLY 4GB to be used/available! With PAE, PA is extended to 64GB. VA is
limited (in both cases) to 4GB.
> But that does only indicate support not use of PAE... I don't see how
> this would be relevant to memtest or the 64-bit kernels anyway.

For x86_64 (good that you are using ONLY these types of kernels) VA is 64
bits, but in reality it is limited to 48 bits (256TB).
> However, cpuid with an x86_64 kernel shows (assumably via the cpuid op

and edx):
> physical address extensions = true

PA supported is 52 bits (1000TBs). So, it is actually limited by number of
DIMM memory slots you have in your system (4 X 8GB DIMM slots). In general
cases up to 256GB, maybe few TB (server configuration wise).
_______
> The current raminit for Nehalem in coreboot is not able to train the two
> 8 GB DIMMs I have tested so far. I have added a debug output to
> choose_reg178 in the first loop before the margins are compared to
> STANDARD_MIN_MARGIN that shows that all margins are 0.

I would suggest to you to do the following: to return to the platform true
BIOS, and to see if BIOS works with your Arrandale i5-520M (Thinkpad 410).
Then you can try several different DIMM configurations, and see if BIOS
supports up to maximum of 32GB (4 X 8GB DIMM).

Then, from these BIOS experiments, you can draw obvious conclusions in
relations with Coreboot MRC algorithm. :-)

Zoran
_______

On Sat, Jan 28, 2017 at 5:41 PM, Stefan Tauner <
[e-mail]> wrote:
> On Sat, 28 Jan 2017 17:01:05 +0100
> Zoran Stojsavljevic  wrote:
> > Hello Stefan,
> >
> > Let me ask you for some other stuff, since I would like to put what I
>
> wrote
> > initially to hold (sleep state, for now).
> >
> > You wrote: *The official specs are not trustworthy IMHO and cpuid(1) and
> > /proc/cpuinfo **show the same physical address width of 36 bits (which
> > would indicate a **maximum of 64 GB).*
> >
> > Question to you: are you dealing with i686 kernel, (32 bit)? It seems to
>
> me
> > that you have Nehalem which complies in IA32 with PAE HW extension, don't
> > you?!
> >
> > This is very important -> *Enabling PAE (by setting bit 5, PAE, of the
> > system register CR4) causes major changes to this scheme...*
>
> Both CPUs have PAE support, yes, but I think I have never even tried to
> boot a 32-bit kernel on them :) However, cpuid with an x86_64 kernel shows
> (assumably via the cpuid op and edx):
>
> physical address extensions            = true
>
> But that does only indicate support not use of PAE... I don't see how
> this would be relevant to memtest or the 64-bit kernels anyway.