RX4640

[Marcel Moolenaar]

On Nov 10, 2004, at 12:17 PM, Arne Schwabe wrote:

> Omar Punzalan <omar.punzalan at ifom-ieo-campus.it> writes:
>
>> I have successfully installed my new (and shiny) freebsd5.3 on one of
>> our hp rx4640 itanium2 system.  With the SMP kernel, it has detected
>> the 4 CPUs correctly.  But then reviewing the /var/log/messages, its
>> sees only 1G of ram:
>>
>> Nov  8 18:10:50 TEST2 kernel: real memory  = 1058357248 (1009 MB)
>> Nov  8 18:10:50 TEST2 kernel: avail memory = 1018814464 (971 MB)
>> Nov  8 18:10:50 TEST2 kernel: FPSWA Revision = 0x10012, Entry =
>> 0xe0000040ffe62050
>> Nov  8 18:10:50 TEST2 kernel: FreeBSD/SMP: Multiprocessor System
>> Detected: 4 CPUs
>>
>> On my web searches, I found out that not even PAE is supported in ia64
>> arch.  I also put hw.physmem=8G inside the loader.conf but no effect.
>> Is there a way for me to convince freebsd that the system has 8Gb of
>> ram?
>
> If I remember PAE was a hack for IA32 to allow more then 4GB but since
> ia64 is 64 bit it should not such a hack for > 4gb

Yes. PAE has nothing to do with ia64. The problem here is that on
some IPF machines (typically the HP rx2600) the physical memory
layout is sparse. It's this sparseness that's causing problems,
because currently we allocate a vm_page array that needs to be
physically contiguous. When a machine has only 2BG of memory,
but 1GB is located above the 4G boundary, we need a vm_page
array that can be larger than the size of any of the usable
memory regions. Not to mention that the vm_page will represent
pages (roughly 3GB worth of it) for memory that isn't even there.

A fix for this is in the pipeline. All it takes is time. If the
memory layout for the machine is not sparse, you can take out
the code from sys/ia64/ia64/machdep.c that skips memory above
4G. Beware of DMA problems caused by the need for bounce buffers
when you do that...

-- 
  Marcel Moolenaar         USPA: A-39004          marcel at xcllnt.net