Reading DDR DCR's within Linux init process

[Sean Parker <supinlick@xxxxxxxxx>]

Hello-

  We're using the PMC-Sierra MSP85XX Sequoia processor... (RM9000 family) The BSP has PMON2000 and Linux 2.6.18 port.

  In configuring/fixing the PMON to properly recognize single- and double-sided RAM modules (no - the BSP did not come from PMC with 'fully' working code in it!) I became familiar with the DDR DCR config process via RAM's I2C access.

  The problem is we want to be able to dynamically read the DDR DCR's (>0xFF000000) setup by PMON, and use the results of that (physical RAM size, i.e. 2G) to call "add_memory_region" in place of the hard-coded 256M "add_memory_region" already called inside the msp85XX-specific "plat_mem_setup" (arch/mips/pmc-sierra/msp85xx/setup.c)

  start_kernel
    ...
    printk(linux_banner)
    setup_arch
      ...
      arch_mem_init
        ...
        plat_mem_setup
          (assign MSP85XX-specific function pointer used later 
                as "late_time_init" below, and assign others)
          add_memory_region( 0x0, 0x10000000, ...) <-- hardcoded 256M
        parse_cmdline_early
          // normal "mem=" interpretation
          add_memory_region  

        // how can I get pointer to 0xFF000000+ DDR DCR here
        // in the process? (before mem_init)

    mem_init
    kmem_cache_init
    setup_per_cpu_pageset
    numa_policy_init
    late_time_init = &py_map_ocd <-- this is where lots of
                                     DCR's are accessed using 
                                     pointers provided by
                                     ioremap - TOO LATE TO CALL
                                     add_memory_region I suppose???
    calibrate_delay
    pidmap_init 
    ...
  <done start_kernel>


  Using command-line parameters (used inside "parse_cmdline_early") is not an option due to our system, and we don't want size-based linux images either.

  I have successfully been able to read the DDR DCRs inside "late_time_init" obtaining a pointer via ioremap call, but that is not a valid call earlier within the context of the setup_arch sequence - is this true? (I get exceptions when I do that) However in late_time_init is too late to call "add_memory_region".

  Does anyone have any ideas? Comments? What other call/procedure  analogous to ioremap can I make earlier to access the DDR DCR insite setup_arch process?

God Bless 
    Sean Parker