Bochs/bochs/docs-html/memory.txt

Memory access handling explanation
by Brendan Trotter:

For  the  local  APIC/s,  any  read  or  write to a CPU's own local APIC is
handled  internally  and  does  not go to the bus. If the read/write misses
this  area then the read/write does go to the bus (where other CPU's ignore
it).

This  means if 2 CPUs have different local APIC addresses and one CPU tries
to  write  to the area used by the second CPU's local APIC, then it will go
to the bus and will not access the second CPU's local APIC.
This applies in all cases (e.g. hyper-threading and dual core work the same).

For  I/O  APICs, the device is on the bus and should override anything that
is  "underneath"  it.  For  example,  if  you  relocate  the  I/O  APIC  to
0x00000000,  then  a  read  or  write  to  this area will not reach the RAM
underneath.  In  a  similar way, if someone maps a PCI device to 0xFEC00000
(or  somewhere  that  overlaps the I/O APIC) then a write to this area will
not reach the PCI device.

This  leads to something like the following for accesses originating from a
CPU:

if (address_is_within_this_CPUs_local_APIC_area)
	do_local_APIC_access();
else if (address_is_within_an_I/O_APIC_area)
	do_I/O_APIC_access();
else if (address_is_within_a_PCI_device_area)
	do_PCI_access();
else if (address_is_within_RAM_area)
	do_RAM_access();
else printf("Bogus address!\n");

For an accesses originating from a PCI device (e.g. PCI bus masters), there
is no access to any CPUs local APIC. It'd go like:

if (address_is_within_an_I/O_APIC_area)
	do_I/O_APIC_access();
else if (address_is_within_a_PCI_device_area)
	do_PCI_access();
else if (address_is_within_RAM_area)
	do_RAM_access();
else printf("Bogus address from PCI device!\n");

In  both  cases  it  is  complicated  by  the configuration of the PCI host
controller/s and any "PCI to PCI" bridges.  Fortunately this can be ignored
by  Bochs as it doesn't support PCI bridges (except for the host controller
itself  which  can  handle all accesses). Bochs may need to worry about the
"PCI  to  LPC"  bridge  though.  For  example, even though a PCI device can
read/write  to  the  I/O  APIC,  an ISA device behind the PCI to LPC bridge
can't. This means for an ISA bus master you'd have something like:

if (address_is_within_a_PCI_device_area)
	do_PCI_access();
else if (address_is_within_RAM_area)
	do_RAM_access();
else printf("Bogus address from PCI device!\n");

This  complicates  things  for the ISA DMA controllers, which should not be
able  to  read/write to the I/O APIC - for e.g. if the I/O APIC base is set
to  0x00000000,  then  an ISA DMA transfer that writes to 0x00000000 should
write to RAM not the I/O APIC (a PCI bus master would write to the I/O APIC
in the same situation).

I'm  not  convinced  modelling  real  hardware  100% correctly is necessary
though  -  it  would only matter for very rare situations (e.g. when the OS
stuffs  things  up  badly).  A normal OS will not stuff things up like this
(i.e.  a  normal  OS won't map the I/O APIC to an area that overlaps RAM or
anything  else).  For  OS  developers  (who  might  stuff  things up), it'd
probably  be  better to panic anyway - e.g. "BX_PANIC: I/O APIC base set to
an address that overlaps RAM or a memory mapped device".

In  general,  the  CPU  has  an "address bus" which consists of data lines,
address  lines  and  2  others  lines. One of these other lines is the "I/O
select"  line  -  if you do "mov [0x000000AA],al" and then do "out 0xAA,al"
you'd  get almost the same thing on the CPUs bus (the only difference would
be the state of the "I/O select" line). When the CPU does an access that is
intended for I/O port space it just asserts the "I/O select" line.

The second line is for SMM which works just like the I/O select line.

When  the  CPU accesses a memory location normally the "SMM select" line is
not asserted and normal memory is accessed. When the CPU is in SMM mode the
"SMM  select" line is asserted for memory accesses. This means that the CPU
can  use  3  completely seperate address spaces (one for normal memory, one
for  I/O  space  and another for SMRAM). How the chipset treats these lines
depends  on  what  the  CPU is used for - for example, these lines could be
ignored  so  that  all types of accesses are the same (which means I/O port
instructions  would  access  memory locations from 0x00000000 to 0x0000FFFF
and  there'd  be no seperate SMRAM area). For "PC compatible" computers the
"I/O  select" line does select a completely seperate address space, but the
"SMM  select" line does not. Instead, the chipset uses it to disable access
to the video display memory (and enable access to the RAM underneath).

Fortunately,  access  to  the SMRAM area is also controlled by the chipset,
such  that  the  CPU can access SMRAM regardless of whether it asserts it's
"SMM select" line or not. As mentioned in my previous email, for the I440FX
chipset  it's called the System Management RAM Control Register (or SMRCR),
and is in the PCI host controller's PCI configuration space at offset 0x72.

Returning  to  what  I  wrote  earlier,  this  leads  to something like the
following for accesses originating from a CPU:

if (address_is_within_this_CPUs_local_APIC_area)
	do_local_APIC_access();
else if ((CPU_is_in_SMM_mode || chipset_SMRCR_enabled) && address_is_within_SMM_area)
	do_SMM_access();
else if (address_is_within_an_I/O_APIC_area)
	do_I/O_APIC_access();
else if (address_is_within_a_PCI_device_area)
	do_PCI_access();
else if (address_is_within_RAM_area)
	do_RAM_access();
else printf("Bogus address!\n");