Re: Include spdmem(4) in GENERIC amd64 & i386 kernels?

[Paul Goyette <paul%whooppee.com@localhost>]

On Mon, 8 Dec 2008, Christoph Egger wrote:

> I will need information about DIMMS, most certainly
> size, physical start address and serial number
> for pmem(9) I'm working on.

Hmmm.  Size of DIMM is available from both dmidecode and spdmem.  None 
of my memory modules have ever shown a valid serial number (this has 
been true for Corsair DDR2, Corsair DDR3, and Patriot DDR3, at least).

Physical address information probably has to be retrieved from the 
memory controller.  I know spdmem/decode-dimms doesn't do this, and a 
cursory trip through dmicode output doesn't show it either (although 
maybe it's referred to by something that doesn't ring any bells).

> Based on this, we can have in the future:
>
> memory offlining, memory hotplugging, xen ballooning,
> persistent dead-page list.

Sounds pretty ambitious - I look forward to seeing this!

> A dead-page list is a list of pages marked as broken/defect.
> Any access to it causes an immediate machine check fault.
> Making it persistent means, the list is kept across reboots
> - given the serial number didn't change.

Hmmm.  I recently had a DDR3 DIMM go bad on me - it kept returning 
apparently random values for bit 0x40.  By random, I mean that it 
returned 0 when it should be 0, 0 when it should be 1, 1 when it should 
be 0, or 1 when it should be 1!  Not predictable, and no correlation to 
the values in adjacent bits.  (This was the Partiot DDR3 which has been 
subesquently replaced with Corsair - no response yet from Patriot on 
what to do to "make good" on their "lifetime warranty".)

I suspect that this type of failure would render the entire DIMM as bad, 
and since this is dual-channel I suspect that it would result in both 
DIMMs being added to the dead-page list, effectively killing the whole 
machine.  :)  At the very least, with dual-channel you would have to 
reconfigure the memory controller to ignore one failed channel, and that 
would mandate a reboot, so persistence would certainly be required.


----------------------------------------------------------------------
|   Paul Goyette   | PGP DSS Key fingerprint: |  E-mail addresses:   |
| Customer Service | FA29 0E3B 35AF E8AE 6651 |  paul%whooppee.com@localhost   |
| Network Engineer | 0786 F758 55DE 53BA 7731 | pgoyette%juniper.net@localhost |
----------------------------------------------------------------------