This simple patch should solve the above problem. Seeing the config, its
seems to be a 32bit box.


It convert tlbstate_lock spin_lock to raw_spin_lock.
Preemption has already been disabled, hence this change shouldn't
affect latency numbers.


Signed-Off-By: Chirag <chirag@linux.vnet.ibm.com>


diff --git a/arch/x86/kernel/tlb_32.c b/arch/x86/kernel/tlb_32.c
index 9bb2363..228849c 100644
--- a/arch/x86/kernel/tlb_32.c
+++ b/arch/x86/kernel/tlb_32.c
@@ -23,7 +23,7 @@ DEFINE_PER_CPU(struct tlb_state, cpu_tlbstate)
 static cpumask_t flush_cpumask;
 static struct mm_struct *flush_mm;
 static unsigned long flush_va;
-static DEFINE_SPINLOCK(tlbstate_lock);
+static DEFINE_RAW_SPINLOCK(tlbstate_lock);
 
 /*
  * We cannot call mmdrop() because we are in interrupt context,



--

[ message continues ]

Chirag,

  Kindly understand what xavier is trying to do. He is Trying to disable smp.
The RT - Linux Code is dependent on the mutex lock to perform some
soft/hard real time processing. This is an problem with the mutex lock.

Kindly check with some one before signing off patches.

Thanks,
Sujit




-- 
--linux(2.4/2.6),bsd(4.5.x+),solaris(2.5+)
--

[ message continues ]

[ Refrain from using "unlisted-recipients" please! ]




The patch is legit.

The problem is simply that functions like flush_tlb_current_task and 
flush_tlb_mm disable preemption and then call flush_tlb_others.
This function locks the tlbstate_lock which is currently a rtmutex. An 
rtmutex can not be called with preemption disabled. The tlbstate_lock is 
static, small and confined. I'm not sure we can change this code (flushing 
the TLB) in a way where we can allow preemption. This may just be a 
latency that we must hit.

-- Steve
--

[ message continues ]