Hi,
I'm currently in the process of integrating some discontig memory support
into the ARM architecture, and I've been looking at the change to MAP_NR()
for to achieve this.
The resulting arithmetic appears to be overly complex for MAP_NR(x):
(((x) - ((x) & const) >> const) +
((&node_data[((x) - const) >> const]->node_mem_map) - cont) / const)
whereas the old contiguous version was:
(((x) - const) >> const)
(I've replaced all the constants with "const" for readability). Obviously
the discontiguous version requires a lot of alu ops to calculate!
Now, MAP_NR is used in essentially one of two ways in the arch/*/mm/* and
mm/* code:
mem_map + MAP_NR(x)
MAP_NR(x) > max_mapnr
What I'd like to propose is to get rid of MAP_NR and instead have a macro
that returned a pointer to the relevent mem_map entry itself. Therefore,
your typical contiguous MAP_ENT(x) macro would be:
(mem_map + (((x) - const) >> const))
ie,
#define MAP_ENT(addr) (mem_map + ((__pa(addr) - PHYS_OFFSET) >> PAGE_SHIFT))
For the discontiguous case, MAP_ENT(x) would be:
((&node_data[((x) - const) >> const]->node_mem_map) +
(((x) & const) >> const))
ie,
#define MAP_ENT(addr) \
(ADDR_TO_MAPBASE((addr)) + (((addr) & (NODE_MAX_MEM_SIZE - 1)) >> PAGE_SHIFT))
This means that we are no longer dependent on the alignment of the node page
maps (I hope - I haven't been through all the references yet), with the exception
of the test with MAP_NR(x) > max_mapnr.
I believe that this test is bogus. If you have the following mem_map
(p = page, h = hole):
<node0 ><node1 ><node2 >
pppphhhhpppphhhhpppp
Surely you want to trap out the holes in this test? If so, then the current code
is going a rather long way about achieving something that is not really effective.
I therefore propose a "MAP_ADDR_VALID(x)" macro.
In the contiguous case, it would simply be:
#define MAP_ADDR_VALID(addr) (MAP_ENT(addr) - mem_map < max_mapnr)
For the discontiguous case:
#define MAP_ADDR_VALID(addr) \
((addr) < NODE_DATA(KVADDR_TO_NID((addr))->node_max_kaddr)
node_max_addr would replace node_start_mapnr, which is only ever initialised and
never subsequently referenced.
I've probably missed something major which throws this idea out, so
comments and/or suggestions welcome. Flames to /dev/null please.
Should this be copied to the linux-mm list as well?
_____
|_____| ------------------------------------------------- ---+---+-
| | Russell King rmk@arm.linux.org.uk --- ---
| | | | http://www.arm.linux.org.uk/~rmk/aboutme.html / / |
| +-+-+ --- -+-
/ | THE developer of ARM Linux |+| /|\
/ | | | --- |
+-+-+ ------------------------------------------------- /\\\ |
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Sat May 06 2000 - 20:03:07 EDT