MCA/INIT: tweaks
Some tweaks to the previous MCA/INIT patch sets.
* Remove the requirement that kernel stacks be aligned on KERNEL_STACK_SIZE.
* Remove the serialization of MCA/INIT handlers returning to SAL. The problem looked like a race but was really caused by a broken prom doing cacheable accesses to the minstate area.
* Print the cpu number and monarch status in the INIT handler.
* Workaround for broken proms that access the minstate area using cacheable addresses.
mca.c | 27 ++++++------------ mca_asm.S | 85 +++++++++++++++++++++++++++++++++++++++++++++++++++------- vmlinux.lds.S | 1 3 files changed, 85 insertions(+), 28 deletions(-)
Index: linux/arch/ia64/kernel/vmlinux.lds.S
=================================--- linux.orig/arch/ia64/kernel/vmlinux.lds.S 2005-09-08 14:06:55.344665850 +1000
+++ linux/arch/ia64/kernel/vmlinux.lds.S 2005-09-08 14:07:32.465527377 +1000
@@ -165,7 +165,6 @@ SECTIONS
__init_end = .;
/* The initial task and kernel stack */
- . = ALIGN(KERNEL_STACK_SIZE);
.data.init_task : AT(ADDR(.data.init_task) - LOAD_OFFSET)
{ *(.data.init_task) }
Index: linux/arch/ia64/kernel/mca.c
=================================--- linux.orig/arch/ia64/kernel/mca.c 2005-09-08 14:06:55.281198033 +1000
+++ linux/arch/ia64/kernel/mca.c 2005-09-08 14:07:32.465527377 +1000
@@ -90,7 +90,6 @@
/* Used by mca_asm.S */
u32 ia64_mca_serialize;
-s32 ia64_mca_init_leave = -1;
DEFINE_PER_CPU(u64, ia64_mca_data); /* = __per_cpu_mca[smp_processor_id()] */ DEFINE_PER_CPU(u64, ia64_mca_per_cpu_pte); /* PTE to map per-CPU area */
DEFINE_PER_CPU(u64, ia64_mca_pal_pte); /* PTE to map PAL code */
@@ -704,11 +703,15 @@ ia64_mca_modify_original_stack(struct pt
msg = "occurred in user space";
goto no_mod;
}
- if ((r12 & -KERNEL_STACK_SIZE) != r13) {
+ if (r13 != sos->prev_IA64_KR_CURRENT) {
+ msg = "inconsistent previous current and r13";
+ goto no_mod;
+ }
+ if ((r12 - r13) >= KERNEL_STACK_SIZE) {
msg = "inconsistent r12 and r13";
goto no_mod;
}
- if ((ar_bspstore & -KERNEL_STACK_SIZE) != r13) {
+ if ((ar_bspstore - r13) >= KERNEL_STACK_SIZE) {
msg = "inconsistent ar.bspstore and r13";
goto no_mod;
}
@@ -717,7 +720,7 @@ ia64_mca_modify_original_stack(struct pt
msg = "old_bspstore is in the wrong region";
goto no_mod;
}
- if ((ar_bsp & -KERNEL_STACK_SIZE) != r13) {
+ if ((ar_bsp - r13) >= KERNEL_STACK_SIZE) {
msg = "inconsistent ar.bsp and r13";
goto no_mod;
}
@@ -907,13 +910,6 @@ ia64_wait_for_slaves(int monarch)
}
}
-static void
-mca_init_leave(int cpu)
-{
- while (cmpxchg_acq(&ia64_mca_init_leave, -1, cpu) != -1)
- cpu_relax();
-}
-
/*
* ia64_mca_handler
*
@@ -968,7 +964,6 @@ ia64_mca_handler(struct pt_regs *regs, s
}
set_curr_task(cpu, previous_current);
- mca_init_leave(cpu);
monarch_cpu = -1;
}
@@ -1218,8 +1213,8 @@ ia64_init_handler(struct pt_regs *regs,
oops_in_progress = 1; /* FIXME: make printk NMI/MCA/INIT safe */
console_loglevel = 15; /* make sure printks make it to console */
- printk(KERN_INFO "Entered OS INIT handler. PSP=%lx\n",
- sos->proc_state_param);
+ printk(KERN_INFO "Entered OS INIT handler. PSP=%lx cpu=%d monarch=%ld\n",
+ sos->proc_state_param, cpu, sos->monarch);
salinfo_log_wakeup(SAL_INFO_TYPE_INIT, NULL, 0, 0);
previous_current = ia64_mca_modify_original_stack(regs, sw, sos, "INIT");
@@ -1230,9 +1225,8 @@ ia64_init_handler(struct pt_regs *regs,
cpu_relax(); /* spin until monarch enters */
while (monarch_cpu != -1)
cpu_relax(); /* spin until monarch leaves */
- printk("slave returning %d\n", cpu);
+ printk("Slave on cpu %d returning to normal service.\n", cpu);
set_curr_task(cpu, previous_current);
- mca_init_leave(cpu);
ia64_mc_info.imi_rendez_checkin[cpu] = IA64_MCA_RENDEZ_CHECKIN_NOTDONE; return;
}
@@ -1271,7 +1265,6 @@ ia64_init_handler(struct pt_regs *regs,
}
printk("\nINIT dump complete. Monarch on cpu %d returning to normal service.\n", cpu); set_curr_task(cpu, previous_current);
- mca_init_leave(cpu);
monarch_cpu = -1;
return;
}
Index: linux/arch/ia64/kernel/mca_asm.S
=================================--- linux.orig/arch/ia64/kernel/mca_asm.S 2005-09-08 14:06:55.272410182 +1000
+++ linux/arch/ia64/kernel/mca_asm.S 2005-09-08 14:07:32.109131175 +1000
@@ -319,11 +319,6 @@ END(ia64_os_mca_virtual_begin)
// release lock
LOAD_PHYSICAL(p0,r3,ia64_mca_serialize);;
st4.rel [r3]=r0
- ;;
- mov r31=-1
- LOAD_PHYSICAL(p0,r3,ia64_mca_init_leave);;
- ;;
- st4.rel [r3]=r31
br b0
@@ -414,11 +409,6 @@ END(ia64_os_init_virtual_begin)
1:
mov b0=r12 // SAL_CHECK return address
- ;;
- mov r31=-1
- LOAD_PHYSICAL(p0,r3,ia64_mca_init_leave);;
- ;;
- st4.rel [r3]=r31
br b0
//EndMain//////////////////////////////////////////////////////////////////////
@@ -626,6 +616,81 @@ ia64_state_save:
;;
st8 [temp1]=temp3 // save ar.lc
+ // FIXME: Some proms are incorrectly accessing the minstate area as
+ // cached data. The C code uses region 6, uncached virtual. Ensure
+ // that there is no cache data lying around for the first 1K of the
+ // minstate area.
+ // Remove this code in September 2006, that gives platforms a year to
+ // fix their proms and get their customers updated.
+
+ add r12*1,r17
+ add r22*2,r17
+ add r32*3,r17
+ add r42*4,r17
+ add r52*5,r17
+ add r62*6,r17
+ add r72*7,r17
+ ;;
+ fc r17
+ fc r1
+ fc r2
+ fc r3
+ fc r4
+ fc r5
+ fc r6
+ fc r7
+ add r172*8,r17
+ add r12*8,r1
+ add r22*8,r2
+ add r32*8,r3
+ add r42*8,r4
+ add r52*8,r5
+ add r62*8,r6
+ add r72*8,r7
+ ;;
+ fc r17
+ fc r1
+ fc r2
+ fc r3
+ fc r4
+ fc r5
+ fc r6
+ fc r7
+ add r172*8,r17
+ add r12*8,r1
+ add r22*8,r2
+ add r32*8,r3
+ add r42*8,r4
+ add r52*8,r5
+ add r62*8,r6
+ add r72*8,r7
+ ;;
+ fc r17
+ fc r1
+ fc r2
+ fc r3
+ fc r4
+ fc r5
+ fc r6
+ fc r7
+ add r172*8,r17
+ add r12*8,r1
+ add r22*8,r2
+ add r32*8,r3
+ add r42*8,r4
+ add r52*8,r5
+ add r62*8,r6
+ add r72*8,r7
+ ;;
+ fc r17
+ fc r1
+ fc r2
+ fc r3
+ fc r4
+ fc r5
+ fc r6
+ fc r7
+
br.sptk b0
//EndStub//////////////////////////////////////////////////////////////////////
Quoted 9 lines · Keith Owens
Some tweaks to the previous MCA/INIT patch sets.
* Remove the requirement that kernel stacks be aligned on
KERNEL_STACK_SIZE.
* Remove the serialization of MCA/INIT handlers returning to SAL. The problem looked like a race but was really caused by a broken prom doing cacheable accesses to the minstate area.
* Print the cpu number and monarch status in the INIT handler.
* Workaround for broken proms that access the minstate area using cacheable addresses.
With the above tweaks, the new MCA/INIT handlers pass all my stress tests. On SGI systems I can send INIT tens of times without any problem, the system dumps the tasks and keeps going. I was also running ia64regcheck at the same time, it passed the test, no registers were corrupted by INIT.
All the problems that have been reported against the new MCA/INIT handlers have been caused by SAL not conforming to the SAL specification. Some versions of SAL only call the monarch cpu and not the slaves for INIT. Some versions of SAL call all cpus as monarchs. Some versions of SAL do not resume correctly after INIT. Even on these broken versions of SAL, the new OS handlers give better results that the existing OS handlers. On working versions of SAL, INIT is now fully recoverable. On working versions of SAL, a recoverable MCA which needs to send INIT in order to rendezvous can now be successfully resumed.
This code is definitely ready for inclusion in 2.6.13-rc1.
Quoted 1 line · Keith Owens
This code is definitely ready for inclusion in 2.6.13-rc1.
Of course that should be 2.6.14-rc1.
Quoted 3 lines · Keith Owens
Quoted 1 line
This code is definitely ready for inclusion in 2.6.13-rc1.
Of course that should be 2.6.14-rc1.
I'm happy that the ia64 parts show a definite improvement on the buggy-SAL systems that I can test on. I'll take your word that this runs fine on SN2. The cleanups to minstate.h alone make this patch beautiful in my eyes :-)
We need to go to linux-kernel to get some blessings for the changes to <linux/sched.h> and kernel/sched.c. The very first question is likely to be: "Why are curr_task() and set_curr_task() exported?" They are only used in mca.c ... which can't be built as a module.
-Tony
Quoted 10 lines · Tony
Quoted 3 lines · Keith Owens
Quoted 1 line
This code is definitely ready for inclusion in 2.6.13-rc1.
Of course that should be 2.6.14-rc1.
I'm happy that the ia64 parts show a definite improvement on the buggy-SAL systems that I can test on. I'll take your word that this runs fine on SN2. The cleanups to
minstate.halone make this patch beautiful in my eyes :-)
Quoted 5 lines
We need to go to linux-kernel to get some blessings for the changes to <
linux/sched.h> andkernel/sched.c. The very first question is likely to be: "Why arecurr_task()andset_curr_task()exported?" They are only used inmca.c... which can't be built as a module.
Tools like kgdb, kprobes and kdb need curr_task() to check if a task is running on a cpu or not, those components can be modular. set_curr_task() does not need to be exported, it is a hang over from some early testing. I'll respin the tweaks patch (number 8) to remove the export of set_curr_task().