INIT handler modification
Dear all,
first of all, my best wishes for this new year.
Second, I just wanted to start a discussion about INIT handler considerations.
Here is the main point, the present code of INIT handler for slave CPUs is reduced
to (file arch/ia64/kernel/mca_asm.S) :
______________________________
GLOBAL_ENTRY(ia64_slave_init_handler)
1: br.sptk 1bEND(ia64_slave_init_handler)
______________________________
I think that when the kernel is deadlocked on another CPU than the monarch one,
the traces may be a bit short when I push on the "DUMP" button to have a complete
diagnosis of my system.
Thus, I propose you to change the following code (file arch/ia64/kernel/mca.c)
______________________________
void __initia64_mca_init(void)
{
ia64_fptr_t *mon_init_ptr = (ia64_fptr_t
*)ia64_monarch_init_handler;
ia64_fptr_t *slave_init_ptr = (ia64_fptr_t *)ia64_slave_init_handler;
.....
______________________________
into
______________________________
void __initia64_mca_init(void)
{
ia64_fptr_t *mon_init_ptr = (ia64_fptr_t
*)ia64_monarch_init_handler;
ia64_fptr_t *slave_init_ptr = (ia64_fptr_t *)ia64_monarch_init_handler;
.....
______________________________
What do you think about it ? Do you fear it may induce an unwanted behavior towards MCA handling ?
Best regards,
Francois WELLENREITER
Quoted 10 lines · Francois Wellenreiter
Here is the main point, the present code of INIT handler for slave CPUs is reduced
to (filearch/ia64/kernel/mca_asm.S) :
______________________________
GLOBAL_ENTRY(ia64_slave_init_handler)
1: br.sptk 1bEND(ia64_slave_init_handler)
______________________________
Quoted 5 lines
I think that when the kernel is deadlocked on another CPU than the monarch one,
the traces may be a bit short when I push on the "DUMP" button to have a complete
diagnosis of my system.
You're right, this is a serious problem that needs to be fixed. Unfortunately, it's a little more complicated than the change you propose, because we need per-CPU areas (or locking) for saving the min-state areas from firmware. I think the recent "per cpu MCA/INIT" changes are heading in this direction, so please take a look at those, try them out, and improve them!
Quoted 7 lines
You're right, this is a serious problem that needs to be fixed. Unfortunately, it's a little more complicated than the change you propose, because we need per-CPU areas (or locking) for saving the min-state areas from firmware. I think the recent "per cpu MCA/INIT" changes are heading in this direction, so please take a look at those, try them out, and improve them!
There should be enough parts of Russ Anderson's changes in the
http://linux-ia64.bkbits.net/linux-ia64-test-2.6.11 tree to give this a try. It should also show up in the next -mm patch (it missed 2.6.10-mm2).
You'd need to add some serialization in ia64_init_handler() and/or init_handler_platform() to make sure that all the cpus didn't stomp all over each other printing their registers, and you'll only want one of them to walk the tasklist providing backtraces.
-Tony
Quoted 9 lines · Francois Wellenreiter
Here is the main point, the present code of INIT handler for slave CPUs is reduced
to (filearch/ia64/kernel/mca_asm.S) :
______________________________
GLOBAL_ENTRY(ia64_slave_init_handler)
1: br.sptk 1bEND(ia64_slave_init_handler)
______________________________
The limitations on slave INIT and MCA handlers are mainly due to the lack of kernel stacks to handle these events. Russ Anderson (SGI) is working on patches that add per cpu stacks for MCA and INIT. Once those patches are in the base kernel, we will be changing the INIT and MCA slave handlers to use the multiple stacks. That will give us real processing for the slave events.
Quoted 8 lines · Keith Owens
Francois Wellenreiter <[email protected]> wrote:
The limitations on slave INIT and MCA handlers are mainly due to the lack of kernel stacks to handle these events. Russ Anderson (SGI) is working on patches that add per cpu stacks for MCA and INIT. Once those patches are in the base kernel, we will be changing the INIT and MCA slave handlers to use the multiple stacks. That will give us real processing for the slave events.
Tony has put the patches in his test-2.6.11 tree. Those patches should have enough support to allow the INIT and MCA slave handlers to use the multiple stacks.
Thanks,
Signature · 2 lines
Russ Anderson, OS RAS/Partitioning Project Lead SGI - Silicon Graphics Inc [email protected]