AGPGART driver for ArticiaS - ioremap() problem

[Benjamin Herrenschmidt]

> Actually at least the bridge does everthing right until now (IMHO :-). The
> bridge is detected and also the aperture size is read out correctly from the
> registers. After that, the agp_backend_initialize() function in
> drivers/cahr/agp/backend.c calls the create_gatt_table() function of the
> driver (articias_create_gatt_table). Based on the detected aperture size and
> its definitions, this function allocates pages for the GATT table
> (alloc_gatt_pages). The function then tries to map this pages with
> ioremap_nocache(virt_to_gatt(table, (PAGE_SIZE * (1 << page_order))). But
> ioremap_nocache() cannot remap the pages, because the following code snipped
> in __ioremap (arch/ppc/mm/pgtable.c) is trigged:

Yes, you are not allowed to ioremap RAM on ppc at least not with
CONFIG_6xx. You should leave it cacheable and use explicit cache flush
like the UniNorth driver.

> As it can be seen in the following log, the table address (0xde200000 ->
> virt_to_gart(table) = 0x1e200000) is lower than the maximum amount of memory
> 0x20000000 (512MB) and therefore __ioremap() does not remap the pages.