Hi Tom,
So I don't know who is supposed to be handling this (Wim?), but the
patch itself looks suspicious.
On Sun, Jan 15, 2012 at 8:02 PM, Maxim Uvarov<maxim.uvarov@xxxxxxxxxx> wrote:- set_memory_x((unsigned long)bios32_entrypoint, (2 * PAGE_SIZE));
+ set_memory_x((unsigned long)bios32_entrypoint& PAGE_MASK, 2);
If it wasn't page-aligned to begin with, then maybe it needs three pages now?
I have been looking at the code again and basically we have for 32 bit the following sequence:
1) scan/search from 0x0f0000 through 0x0fffff, inclusive (in steps of 16 bytes) until we find
the 32-bit BIOS Service Directory with signature == PCI_BIOS32_SD_VALUE (=0x5F32335F ="_32_").
2) If we find this area then we first do a checksum check to see if it's a valid area.
3) if it's a valid area then we will check this area for a $CRU record.
the code for this is as follows:
/*
* According to the spec, we're looking for the
* first 4KB-aligned address below the entrypoint
* listed in the header. The Service Directory code
* is guaranteed to occupy no more than 2 4KB pages.
*/
map_entry = bios_32_ptr->entry_point& ~(PAGE_SIZE - 1);
map_offset = bios_32_ptr->entry_point - map_entry;
bios32_map = ioremap(map_entry, (2 * PAGE_SIZE));
if (bios32_map == NULL)
return -ENODEV;
bios32_entrypoint = bios32_map + map_offset;
cmn_regs.u1.reax = CRU_BIOS_SIGNATURE_VALUE;
set_memory_x((unsigned long)bios32_entrypoint, (2 * PAGE_SIZE));
asminline_call(&cmn_regs, bios32_entrypoint);
=> So if I understand it correctly then map_entry is page aligned. And thus bios32_map is also page aligned.
Wouldn't it then not make more sense to do a:
set_memory_x((unsigned long)bios32_map, 2);
- set_memory_x((unsigned long)cru_rom_addr, cru_length);
+ set_memory_x((unsigned long)cru_rom_addr& PAGE_MASK, cru_length>> PAGE_SHIFT);
Same here. If we align the start address down, we should fix up the
length. And should we not align the number of pages up?
In general, a "start/length" conversion to a "page/nr" model needs to be roughly
len += start& ~PAGE_MASK;
start&= PAGE_MASK;
nr_pages = (len + PAGE_SIZE - 1)>> PAGE_SHIFT;
to do things right. But I don't know where those magic numbers come
from. Maybe the "2" is already due to the code possibly traversing a
page boundary, and has already been fixed up. Somebody who knows the
driver and the requirements should take a look at this.
4) if we then found the $CRU record then we do:
physical_bios_base = cmn_regs.u2.rebx;
physical_bios_offset = cmn_regs.u4.redx;
cru_length = cmn_regs.u3.recx;
cru_physical_address = physical_bios_base + physical_bios_offset;
/* If the values look OK, then map it in. */
if (cru_physical_address) {
cru_rom_addr = ioremap(cru_physical_address, cru_length);
if (cru_rom_addr) {
set_memory_x((unsigned long)cru_rom_addr, cru_length);
retval = 0;
}
}
=> Which means that cru_physical_address and cru_rom_addr are not page-aligned.
So if we follow the conversion model that Linus described we get:
set_memory_x((unsigned long)cru_rom_addr& PAGE_MASK,
(cru_length + PAGE_SIZE - 1)>> PAGE_SHIFT);
Can you check this?
Kind regards,
Wim.