sys/doc/memory-test.doc

Message-Id: <199403270118.RAA00584@corbin.Root.COM>
From: David Greenman <davidg@root.com>
To: "Jordan K. Hubbard" <jkh@whisker.hubbard.ie>
Cc: csgr@alpha.ru.ac.za, freebsd-hackers@freefall.cdrom.com
Subject: Re: Boune Buffers : Crash and Burn ^&%$&&^%
Date: Sat, 26 Mar 1994 17:18:04 -0800

>Ummm.  I suppose it would be stupid to wonder whether or not this is good
>memory you just stuck in there? :-)  It does raise the point:  David - what
>does your memory test do when it finds memory errors?

   The order of events is like this:

1) Determine amount of memory by looking at CMOS.
2) Do memory test of this memory, starting at the highest address and going
backwards.
3) If a bad page is found, TRUNCATE the physical memory to end just before
this page. Continue until all memory is tested.

   So all memory past the lowest bad page is ignored. The number of pages that
were chopped are then reported as # of "bad pages" when the memory information
is printed. If there aren't any, then nothing is printed.
   The rationale for not constructing a bit mask and weeding out just the
pages that test bad is a flimsy one - doing so would have required moving the
memory test into the machine independant portion of the VM system, and would
have required much trickier coding. Since the main purpose of this was to
verify that what the CMOS was telling us was correct, and also to identify
when there is bad memory, trying to make the best of it when there really is
some bad memory didn't seem important to me. ..but hey, if somebody wants to
recode it, I won't complain. :-)
   The problem that Terry eluded to, but didn't state, is where you test memory
for the purpose of determining how much you have. This is problematic because
many memory controllers will generate a fatal NMI when you 'test' non-existant
memory (i.e. the parity check fails, and the condition is handled less than
gracefully), and can also happen with some memory controllers that do strange
address wrapping in the non-existant memory area. Compaq machines are alledged
to be suseptable to one of these two problems.

-DG