Re: [kernel-reviewers] a small code review (2414483) Automated g4 rollback of changelist 2396062.

Re: [kernel-reviewers] a small code review (2414483) Automated g4 rollback of changelist 2396062.

[Martin Bligh]

Tim Hockin wrote:
> On Wed, Apr 26, 2006 at 10:12:14AM -0700, Ken Harrenstien wrote:
> 
>>That doesn't work because IIRC it only reports the amount of memory
>>the kernel has been told (eg via "mem=") to manage in a certain sense,
>>not how much is actually physically available.
>>
>>The I2 netboot kernel would really REALLY like some exported /proc
>>values that accurately report physical memory (if nothing else, the
>>number of DIMMs and their sizes).  It has to figure this out in order
>>to install the proper kernel with proper LILO command-line args.
> 
> 
> The kernel can't really know how much memory is in the system without
> getting chipset-specific.
> 
> MTRR is a good way to hazard a guess, and will probably be right, but as
> you indicated, BIOS vendors have historically been REALLY bad about
> MTRRs.  Better now, but bad a few years ago.
> 
> SMBIOS (on our boards) *does* accurately report the number of DIMMS and
> their sizes (and more!).  But it only works on Google BIOS.

cc: linux-kernel
bcc: kernel-reviewers

Are you saying our e820 maps and srat tables are wrong? that's a little
worrying ...

M.

Re: [kernel-reviewers] a small code review (2414483) Automated g4 rollback of changelist 2396062.

[Tim Hockin]

On Wed, Apr 26, 2006 at 10:36:54AM -0700, Martin Bligh wrote:
> Tim Hockin wrote:
> >On Wed, Apr 26, 2006 at 10:12:14AM -0700, Ken Harrenstien wrote:
> >
> >>That doesn't work because IIRC it only reports the amount of memory
> >>the kernel has been told (eg via "mem=") to manage in a certain sense,
> >>not how much is actually physically available.
> >>
> >>The I2 netboot kernel would really REALLY like some exported /proc
> >>values that accurately report physical memory (if nothing else, the
> >>number of DIMMs and their sizes).  It has to figure this out in order
> >>to install the proper kernel with proper LILO command-line args.
> >
> >
> >The kernel can't really know how much memory is in the system without
> >getting chipset-specific.
> >
> >MTRR is a good way to hazard a guess, and will probably be right, but as
> >you indicated, BIOS vendors have historically been REALLY bad about
> >MTRRs.  Better now, but bad a few years ago.
> >
> >SMBIOS (on our boards) *does* accurately report the number of DIMMS and
> >their sizes (and more!).  But it only works on Google BIOS.
> 
> Are you saying our e820 maps and srat tables are wrong? that's a little
> worrying ...

e820 is fine, except there's not a good kernel interface to it.  Parsing
dmesg is *always* a last resort.

We don;t currently have SRAT. :)

Re: [kernel-reviewers] a small code review (2414483) Automated g4 rollback of changelist 2396062.

[Andi Kleen]

Martin Bligh <mbligh@google.com> writes:

> Tim Hockin wrote:
> > On Wed, Apr 26, 2006 at 10:12:14AM -0700, Ken Harrenstien wrote:
> >
> >>That doesn't work because IIRC it only reports the amount of memory
> >>the kernel has been told (eg via "mem=") to manage in a certain sense,
> >>not how much is actually physically available.
> >>
> >>The I2 netboot kernel would really REALLY like some exported /proc
> >>values that accurately report physical memory (if nothing else, the
> >>number of DIMMs and their sizes).  It has to figure this out in order
> >>to install the proper kernel with proper LILO command-line args.
> > The kernel can't really know how much memory is in the system without
> > getting chipset-specific.
> > MTRR is a good way to hazard a guess, and will probably be right,
> > but as
> > you indicated, BIOS vendors have historically been REALLY bad about
> > MTRRs.  Better now, but bad a few years ago.
> > SMBIOS (on our boards) *does* accurately report the number of DIMMS
> > and
> > their sizes (and more!).  But it only works on Google BIOS.

Actually quite a lot of SMBIOS report DIMMs more or less correct.
It works even on my laptop.

I implemented support for it in the latest mcelog (--dmi). Not all
and a few big names screw it up (in particular the mappings from
address range to interleave set). But overall it looks surprisingly
good so far. 

-Andi