: load_mmu for SMP systems

In SMP systems, each CPU needs to set up "current_cpu_data.tlbsize". Some CPUs do this initialization in cpu_probe, which is called both by init_arch and start_secondary. However, some CPUs do this in their TLB setup code, which is called via load_mmu. The SMP boot code doesn't currently call load_mmu() for the secondary CPUs. Here's a simple fix for the 2.4 tree.

TLB flush routines that have loops running up to tlbsize will lose if it's not set properly on all CPUs!

Kip

Index: arch/mips/kernel/smp.c
===================================================================
RCS file: /home/cvs/linux/arch/mips/kernel/smp.c,v retrieving revision 1.10.2.24

diff -u -r1.10.2.24 smp.c
@@ -99,6 +99,7 @@
 	unsigned int cpu = smp_processor_id();
 
 	cpu_probe();
+	load_mmu();
 	prom_init_secondary();
 	per_cpu_trap_init();
 

Index: arch/mips64/kernel/smp.c
===================================================================
RCS file: /home/cvs/linux/arch/mips64/kernel/smp.c,v retrieving revision 1.26.2.22

diff -u -r1.26.2.22 smp.c
@@ -103,6 +103,7 @@
 	unsigned int cpu = smp_processor_id();
 
 	cpu_probe();
+	load_mmu();
 	prom_init_secondary();
 	per_cpu_trap_init();
 

OK, I'll add some declarations make it tidier.

Kip

Index: arch/mips/kernel/smp.c
===================================================================
RCS file: /home/cvs/linux/arch/mips/kernel/smp.c,v retrieving revision 1.10.2.24

diff -u -r1.10.2.24 smp.c
@@ -39,6 +39,9 @@
 #include <asm/mmu_context.h>
 #include <asm/smp.h>
 
+extern void load_mmu(void);
+extern void per_cpu_trap_init(void);
+
 /* The 'big kernel lock' */
 spinlock_t kernel_flag __cacheline_aligned_in_smp = SPIN_LOCK_UNLOCKED;
 int smp_threads_ready;	/* Not used */

@@ -99,6 +102,7 @@
 	unsigned int cpu = smp_processor_id();
 
 	cpu_probe();
+	load_mmu();
 	prom_init_secondary();
 	per_cpu_trap_init();
 

Index: arch/mips64/kernel/smp.c
===================================================================
RCS file: /home/cvs/linux/arch/mips64/kernel/smp.c,v retrieving revision 1.26.2.22

diff -u -r1.26.2.22 smp.c
@@ -39,6 +39,9 @@
 #include <asm/mmu_context.h>
 #include <asm/smp.h>
 
+extern void load_mmu(void);
+extern void per_cpu_trap_init(void);      
+
 /* The 'big kernel lock' */
 spinlock_t kernel_flag __cacheline_aligned_in_smp = SPIN_LOCK_UNLOCKED;
 int smp_threads_ready;	/* Not used */

@@ -103,6 +106,7 @@
 	unsigned int cpu = smp_processor_id();
 
 	cpu_probe();
+	load_mmu();
 	prom_init_secondary();
 	per_cpu_trap_init();
 

I instead changed cpu-probe to set tlbsize properly. Nothing wrong with your patch, it just fits better into my Grand Plan (TM) :-)

Yeah, they're going to be sort of slow. There must be a reason for all those GHz processors ;-)

Ralf

Um, it was worse than that if (for example) a complete TLB flush has a "for (i=0; i<0; i++)" loop around it. My board was experiencing occasional userland segfaults thanks to bogus TLB flushing.

Kip